We have asked bot makers who apply for credits if they want to publicly share what they plan to do.  The survey question asks, "Describe your bot idea in 1-3 paragraphs". We collect this information to help other bots target novel ideas to try. Bot makers often change plans as they get feedback, though these descriptions should still be useful snapshots into what the community is doing. See if you can create a bot that tries something others have not done, or is a spin-off of another idea. If you see potential synergy between your bot and another participant, Metaculus may be able to connect you (please reach out to ben \[at] metaculus \[.com]).

## Fall 2026

**4prophets-bot**

> 4prophets is a four-vendor ensemble (GPT, Claude, Gemini, Grok; 2–3 samples each) aggregated by plain median with capped tails — deliberately boring at the aggregation layer, since stackers, judge LLMs, and post-hoc calibration have been repeatedly tested and rejected by open-source competitors. The exploratory bet is upstream: a timestamped evidence store. Every research item is stored with its publication date and tagged to a numbered uncertainty, and forecasters receive a brief organized into explicit time layers (historical base rates / current state / recent momentum) plus a mandatory CONFLICTS section surfacing disagreeing sources, aimed at the briefing-induced herding failure mode several bot makers have reported, where a shared high-quality brief collapses ensemble disagreement. Research is a parallel fan-out plus a document harness that renders JS pages and parses PDFs/CSVs, since official sources often live in formats news APIs don't carry. On high ensemble disagreement, the bot runs one targeted extra research round and re-forecasts with the full ensemble: extra research rather than a judge, following the BTF-2 and AIA Forecaster findings. The timestamp design also gives leak-controlled pastcasting for free: any resolved question can be replayed using only evidence dated before the forecast.

**astra\_edge\_bot**

> I’m developing one cost-bounded forecasting bot that focuses on interpreting resolution criteria precisely and checking whether its evidence is current. For each question, the planned pipeline records the resolution conditions, relevant base rates, dated sources, evidence for and against the outcome, and unresolved uncertainties. It then produces a probability distribution and concise explanation, with automated checks for valid output and consistent probabilities. The initial experiment will test whether explicit checks of resolution wording and evidence freshness improve forecasting compared with a simpler baseline under the same inference budget. This is a proposed approach, not a claim of novelty or demonstrated performance. Local synthetic format checks are complete; live integration and prospective evaluation remain pending. Development will use synthetic or closed historical questions, with historical results treated cautiously because of possible model knowledge contamination. Once participation is approved, the bot will forecast autonomously without human editing or selective reruns of live answers, while tracking forecast quality, coverage, failures, and cost per question.

**AUREUS200**

> My bot will use a cost-aware, uncertainty-driven forecasting pipeline rather than applying the same expensive reasoning process to every question. It will first generate a structured baseline forecast using lightweight reasoning, explicit decomposition of the question, and evidence gathering. It will then estimate uncertainty and disagreement across candidate forecasts. Only questions with high uncertainty, conflicting evidence, or unusually high expected information value will trigger deeper research, additional model calls, or alternative reasoning paths. The research goal is to test whether selective escalation can improve forecasting accuracy and calibration per dollar of inference cost. I also plan to compare single-model forecasts against small ensembles and disagreement-based aggregation, while tracking calibration, error patterns, source diversity, and marginal value of additional computation. The system will remain fully automated and will record enough metadata to evaluate which reasoning strategies actually improve forecasts rather than simply consuming more compute.

**delphi**

> delphi is a web-grounded forecasting agent built on a frontier reasoning LLM. For each question it runs live retrieval keyed on the question text and resolution criteria, pulls the top results, fetches and strips the leading source pages, and feeds that evidence to the model with an explicit "Retrieved sources" citation list, so every forecast is grounded in current information rather than parametric memory. Research is cached briefly so repeated runs don't re-search, and the pipeline fails soft to reasoning-only if retrieval is unavailable. The design goal is calibration under sparse, non-linguistic uncertainty rather than raw scale: a reasoning model does the probability judgment, grounded search supplies the facts, and the emphasis is on faithful evidence handling (dated sources, explicit resolution-criteria reasoning). With OpenRouter credits I'd run and ensemble stronger SOTA models (GPT-5.x / Claude / Gemini) via OpenRouter, add AskNews for higher-quality retrieval, and build a pastcasting harness to tune calibration on already-resolved questions before forecasting live ones.

**devinjones-bot**

> An ensemble forecaster built strictly from the published FutureEval evidence base (Fall 2025 survey, calibration-adjustment analysis). Each question gets a two-round native-web-search research brief, then five independent forecasts across three frontier model families, aggregated in code: median, tail caps, and a triggered second-look check on any extreme aggregate to guard against the known already-resolved failure mode. Continuous questions use a deterministic percentile-to-CDF pipeline with the API's monotonicity constraints satisfied by construction. The emphasis is reliability engineering: per-question failure containment, duplicate detection via the post detail endpoint, and a full JSONL audit trail of every run — which doubles as the training set for Platt-scaling recalibration once enough forecasts resolve.

**dinghyx3-bot**

> dinghyx3-bot is a calibration-first ensemble forecaster. The design comes out of several months running automated prediction-market trading systems against Kalshi, where the recurring lesson was that overconfidence is what costs you and that most apparent edges dissolve once you measure them honestly. In that work calibration was a compromise — the only signal with enough statistical power at my sample sizes, and I was careful never to treat it as proof of profitability. Here it isn't a compromise: the tournament scores probabilistic accuracy against other forecasters rather than hit rate, so calibration is the actual objective rather than a stand-in for it. That makes this a better-posed problem than the one I came from, and the habits transfer directly. I'm not attempting anything architecturally novel. The stated finding that scaffolding rather than model choice is the current differentiator is worth taking literally, so I plan to start from the official template, read the open-source bots from past winners before writing much of my own, and only diverge where I have a specific reason. Structurally it's a two-stage scaffold built around a hard constraint: with AskNews capped at \~1k calls/month, a bot gets roughly one research pass per question, so retrieval has to be made to count rather than iterated. A cheap model condenses retrieved news into a structured evidence summary — base rates, what would have to change, what's already priced in — and then two or three frontier models from different providers forecast independently from that same brief. Their probabilities are aggregated rather than argued to consensus, since independent errors partially cancel and ensembling is a more reliable calibration gain than any single model choice. Every forecast emits a written rationale naming its base rate and the specific evidence that moved it off that base rate. The concrete edge I'm targeting is non-binary questions, where published results indicate the pro-versus-bot gap is widest. Pulling the Summer 2026 question set via the API, binary is only about half of it — roughly 23% numeric, 16% multiple choice, 11% discrete — so half the tournament sits in exactly the formats bots handle worst. That gap looks more like a coherence failure than a knowledge failure: models reason sensibly about each option or bucket and then emit distributions that are individually plausible but jointly incoherent. Enforcing coherent normalization across options, and shrinking toward a sensible prior when the ensemble disagrees, is a calibration problem rather than a research problem — which is the kind I have tooling and habits for. The rest is uptime: coverage multiplies the score, so the harness is built to run unattended with health gates and failure alerting, the same pattern my trading systems already use.

**emreturan\_bot**

> A multi-stage pipeline built on a decorrelated ensemble: several frontier models (Claude, GPT, Gemini, plus Grok as a low-correlation member) run under deliberately different prompts - some reasoning from reference classes and base rates (outside view), others from causal mechanism (inside view) - aggregated with a median/trimmed combination. Two modules target known weak spots: a reference-class engine that retrieves analogous resolved questions and historical base rates as structured priors, and a numeric-question calibration layer (PIT / location-scale) for the continuous-question miscalibration most bots handle poorly. An adversarial self-critique pass runs before final aggregation, outputs are clipped to a sane band with a resolved/unresolved guard, and research is agentic and multi-source with a relevance filter. We prioritize full question coverage and submission reliability, iterating via an offline harness on resolved questions plus MiniBench.

**FutureSpark-Bot**

> Most bots send every question to an LLM. Mine doesn't. If a question asks whether a temperature or an economic series will go up, I answer it with an actual weather forecast or the series' own history, with no LLM involved. If a question tracks a prediction market, I start from the current market price. The LLM only handles questions where neither of those applies. Each part has to prove it helps before it goes live. A new component only gets switched on if it beats what's already running by more than two standard errors out of sample, and otherwise it falls back to the old behavior. Several ideas I was confident about got rejected by that test. The reason for the caution is that questions in the same round often aren't independent. Forty-eight weather stations in one round share the same weather system, so you effectively have about three independent data points, not forty-eight. Measuring error across rounds instead of across questions makes a lot of promising-looking signals disappear.

**killian**

> Ora (Killian's engine) is a pipeline forecaster built to be measured, not just to submit. It starts by reading the resolution criteria carefully — fetching the resolver page and linked sources to work out exactly what would count — before any research happens. Research is a multi-source fan-out (news retrieval plus web search) compiled into a structured dossier. A cross-family ensemble of frontier models (Anthropic, OpenAI, Google) then forecasts, with each seat reasoning from both an outside view (reference class, base rates, status quo) and an inside view (causal model, scenarios), and the ensemble is aggregated into the final probability; numeric questions get a full CDF submission. The bot is instrumented end to end: every seat answer, every dossier, and every dollar of cost is stored, so improvements are validated by replaying resolved questions offline rather than by fielding multiple live bots. Summer results (joined 10 July, mid-season): 208 of 208 open submission windows hit, currently rank 29 of 268 with 138 questions scored (peer sum +1211.9) and 53 still awaiting resolution. On a per-question basis that's a +14.5 average peer score — 16th of the 173 bots with fifty or more scored questions, against +18.2 for the tournament leader. The Fall workstream is calibration: fitting corrections from our own resolved history as the sample grows. Cost is engineered down every month — \$3.51 per question in July, \$2.28 on the current pipeline — through itemized per-call cost tracking, caching, and continuous tuning, without dropping the full frontier panel.

**Panshul42**

> I'd like to leverage the best frontier models available, and build a minimal design with lots of tools around a singular agent, rather than the 'swarm of agents' approach I did for previous tournaments. That seems to be yielding better results.

**randomwalk-bot**

> Multi-sample ensembling (5+ samples per question, median aggregation), a research loop over news retrieval before elicitation, and separate calibration handling for binary / multiple-choice / numeric / date question types. I work in quantitative finance and intend to apply proper scoring-rule discipline (pre-registered prompts, no post-hoc tuning mid-tournament).

**seergiii-bot**

> Most template-derived bots follow a single pipeline: retrieve news, hand it to an LLM, ask for a probability. My hypothesis is that this ordering is the main source of miscalibration, because the model sees the most recent and most dramatic evidence before it has any anchor, and over-updates on it. This shows up as overconfidence on questions where the status quo is overwhelmingly likely to hold. My bot inverts the order. Stage one runs before any retrieval: the model is asked to construct an explicit reference class for the question and state a base rate from that class alone, with no news access. Stage two retrieves evidence and evaluates it against a fixed likelihood-ratio rubric, so the update is bounded and has to be justified in odds terms rather than produced free-hand. The final forecast is the prior updated by the capped likelihood ratio, not a number the model emits in one pass. The measurable claim is that separating prior formation from evidence updating improves calibration on low-probability and status-quo questions, at some cost on questions where genuinely new information dominates. I plan to test this by pastcasting both variants over resolved questions and comparing Brier scores and calibration curves, then submitting the winning configuration to MiniBench. The credits are needed because this design costs roughly twice a single-pass bot per question, and the pastcasting evaluation is where most of the spend goes.

**Signal2Noise**

> Signal 2 Noise is an independent AI forecasting engine competing in FutureEval and MiniBench. As of August 2026, it uses GPT-5.5 with live Exa web retrieval to forecast binary, multiple-choice and numeric questions. For harder questions, it breaks the problem into smaller questions, retrieves evidence for each, and combines the results. Date and discrete-outcome questions use a separate pipeline based on empirical base rates. We have made 12,000 forecasts across 34 domains, including geopolitics and conflict, politics and law, science and health, markets, and AI, covering five question formats. The project is developed in public: an engineering log at signal2noise.uk documents the methodology and failures, while every forecast is recorded with its full provenance. Development is evidence-led, using a leak-safe frozen-corpus backtest, resolved-question scorecards, and offline comparisons with the crowd aggregate. Each forecast costs roughly \$0.02, allowing additional credits to be used for a stronger base model, more samples, or deeper analysis of difficult questions.

**whiskeyjack-bot**

> A major focus that keeps coming up in my work is source/claim validation. It can be easy for a human with expert knowledge in a domain to intuitively know who is a trustworthy source and who is not, but it is a trickier interdisciplinary problem to automate and quantify that process. The focus of my bot is the attribution instrument; what information was available and considered at the time of the forecast? This can then be tracked and sources with a better record can be given a higher weight. There's quite a few parallels with Metaculus and this; contrarian forecasts with a longer lead time should be more highly valued. There's risks about overfitting here, 'consensus opinion', and it will likely be murky for geopolitical questions. Solving those issues is half the fun though!

**ROBO**

> The bot starts as the Metaculus template bot, unmodified, and every change to it is a single knob whose default reproduces the template exactly. The calibration step, for example, ships inert, and a test asserts that with it off the bot returns the parent's own output object rather than a rebuilt copy. The point is attribution: when a season's score moves I want to know which change moved it, instead of having rewritten five things at once and guessing. The lever I want to test first is not a prompt. The prize take is the square of the summed peer score, and a question left unforecast contributes zero, so coverage is a scoring lever rather than plumbing. Questions launch at random hours and stay open briefly, so the practical question is how reliably a runner catches them. I plan to measure my own miss rate as a first-class metric alongside accuracy, and to treat a missed question as a scored failure rather than an operational hiccup. The third lever is calibration. The log score punishes a confident error far harder than it rewards a confident correct answer, so the question worth testing is whether shrinking the aggregate toward a base rate pays on peer score, and by how much. I will test that on closed questions only. What I am deliberately not doing is finetuning, or prompt archaeology on one model: the arms are structural, and I would rather report a clean null on three of them than an unattributable gain.

**ForecastForge-bot**

> My bot will be able to produce predictions using information retrieval and probabilistic reasoning techniques. It will collect evidence for each question and determine which key drivers and uncertainties exist and will generate a calibrated probability instead of just going off-instinct or an impressionistic model response. Using MiniBench results, I will be making sure to gradually tweak the process of forecasting, particularly the calibration, the evidence and how the Bot will deal with conflicting evidence presented. The aim is to develop a lightweight automatic forecasting system, which will be consistently operating under various domains and need to improve its performance according to the performance measured during training.

**Oliveira-Vantage**

> Oliveira-Vantage is a student-led forecasting research project built around auditable ex-ante prediction. In our electoral work, forecasts are frozen before the event with cryptographically verifiable timestamps and evaluated against subsequent outcomes. We want to bring the same discipline to FutureEval while testing a structured forecasting architecture around strong frontier models. Our main differentiator is a semantic-first forecasting layer. Before researching or assigning a probability, the bot explicitly resolves what the question is actually asking: target event or state, resolution criteria, time horizon, entity, units and denominator, required observable, and distinctions such as intention vs action, approval vs implementation, capability vs event, state vs transition and indicator vs outcome. Only after this question contract is resolved does the bot perform research, establish a status-quo/base-rate anchor, generate multiple independent forecasts, reconcile disagreement and produce a calibrated probability. Development will use strict-cutoff pastcasting, MiniBench, community comparison, error logging and calibration analysis. The seasonal bot itself will operate autonomously; human review is used offline to improve later versions rather than to alter individual forecasts.

**Vrchcab**

> Vrchcab is an automated forecasting bot that uses an LLM to analyze forecasting questions, gather relevant information, consider multiple possible outcomes, and produce calibrated probability estimates. I plan to start with a relatively simple and lightweight system and improve it based on its forecasting performance. I want to test whether combining structured reasoning, current information, and comparison with independent forecasts can improve calibration. I will evaluate the bot on past and current questions and adjust the approach during the tournament rather than relying on a fixed prompt.

**Akela**

> Akela is a custom bot, not a fork of the template. It runs in two stages. First a single research call with live web search builds a briefing for the question. Then five separate LLM members each produce a full distribution from that briefing independently, and those five distributions are aggregated into the submitted forecast. Aggregating at the distribution level rather than asking one model for a final number is deliberate: the spread between members is itself information, and I use it to decide which forecasts to review by hand. The part I think is less common is that the bot is a branch of a larger forecasting system rather than a standalone tournament entry. Its parent project is a Finnish-language public-economy dashboard that forecasts official statistics releases, publishes those forecasts before the release, and then scores itself against what actually happened. The same ensemble machinery, the same cost ledger, and the same learning loop serve both. When a forecast resolves in either domain, the system writes a short lesson about what the reasoning missed, and those lessons are retrieved into later prompts for related questions. FutureEval is the independent scoreboard for machinery that otherwise only grades itself. On the engineering side I lean on two things that keep the cost per question low enough to run a full season: batch submission when a question closes more than six hours out, which halves the model price, and prompt cache warming, where the first ensemble member runs alone so the remaining four read the shared context from cache instead of each writing their own copy. Measured on a practice round, the warming change moved cache hit rate from 53 percent to about 90 percent.

**edisonymy-bot**

> edisonymy-bot is a skill-only bot: all forecasting judgment lives in a pair of plain-markdown Agent Skills (forecast, calibrate) plus a small stdlib-only Python core for the arithmetic (journal, scoring, pooling, CDF construction). The tournament harness is a thin consumer that runs the same forecast skill headlessly through Claude Code with web search; no forecasting logic lives in the bot. The method is readable and forkable, model-agnostic, and identical to what a human can invoke in a chat. The full track record is an append-only public JSONL journal committed on every run, with a per-question trace of what each agent call thought. Design for Fall 2026 (v0.4.28): a cheap triage call sets effort; medium/high questions get 3–4 fully independent research runs (each must consult at least 3 distinct sources), pooled by geometric-mean odds for binaries, quantile averaging for continuous questions, and per-option geometric mean for multiple choice. A supervisor call then sees every run's evidence and reasoning, lists the disagreements, classifies each as factual or judgment, settles the factual ones, and issues the final number without averaging; its web-research budget is conditional on how much the runs disagree. Continuous questions submit five percentiles with explicit tail escape masses, built into a monotone CDF. The bot never sees any crowd aggregate. Every change ships as a preregistered paired test scored at resolution. Summer 2026 record. MiniBench, all four waves entered: 12th, 15th, 12th, and 5th of 182 (+944 points) in the 2026-08-24 wave, three of four in the money; 9th of 112 on the four-wave total and the top-scoring source-available bot, behind only two Metaculus baseline bots. Seasonal FutureEval: entered 2026-07-06 and finished 95th of 282 on total points, dragged down by six July numeric questions that resolved outside the submitted range. After the August fixes the rate was about +17 points per question on 41 questions, which would sit in the top six among bots with comparable coverage, but the sample is small and the fix–effect attribution is not clean. Fall, with full coverage from the start, is the real test. Fall changes target the measured deficit, continuous questions: pooled quantiles across runs, empirical-dispersion checks in research, and a preregistered CDF-smoothing test.

**Marty-Bot**

> I am building an autonomous general-purpose forecasting bot designed around rigorous research, probabilistic calibration, and systematic evaluation. The system will use multiple independent research and forecasting perspectives, structured evidence from high-quality sources where available, and a robust aggregation layer to produce forecasts across binary, multiple-choice, and numeric questions. A major focus of the project is empirical iteration: comparing architectural changes against a strong baseline through prospective testing, tracking recurring failure modes, and improving only components that demonstrate measurable forecasting gains. My goal is to build a highly competitive FutureEval system while also learning which AI forecasting techniques genuinely generalize.

**muller-ganhador-bot**

> My bot is deliberately small. It is built on forecasting-tools and overrides only the research step and the three forecast methods. I made that choice after reading the Spring 2026 bot maker survey, where the loudest recurring theme was that added complexity hurt: one maker fell from 20th to 63rd after adding logit-space aggregation, extremization and crowd-blending, and several others described tearing out elaborate architectures. Rather than build another scaffold, I want to spend the season measuring a few things properly. The thing I actually want to test is the contested claim about ensembling. Schneider and Schramm found that deliberation between different models improved Brier from 0.162 to 0.153 on 202 questions from this tournament, while three instances of the same model changed nothing. But the Spring 2026 correlation table scores aggregating multiple forecasts at r \= -0.19, and Metaculus states plainly that no Spring result survives Benjamini-Hochberg correction. Those point in opposite directions, and as far as I can tell nobody has run a clean within-season A/B on it. So I plan to submit two bots that differ only in that switch: one forecasting with a heterogeneous three-model ensemble aggregated by trimmed mean, one with a single model, everything else identical and fixed before the season opens. The second thing is calibration. Platt scaling has real significance behind it while hand-tuned extremization scores r \= -0.30, so I fit the two coefficients from my own resolved questions rather than guessing them, and they ship disabled until there is enough data to fit. I also instrument failure and not only accuracy: the bot raises an alarm on any forecast that lands on exactly 0.5, after reading that a Spring participant lost a season to a silent fallback doing exactly that.

**ORACLE-X**

> ORACLE-X is an open-source autonomous forecasting bot focused on producing well-calibrated probabilistic forecasts rather than relying on a single LLM judgment. The system is being designed around a pipeline that separates research, forecasting, validation, and final probability aggregation. It will use current information when appropriate, historical base rates, scenario analysis, and independent model estimates before producing a final forecast. A key focus of ORACLE-X will be reducing common LLM forecasting failures such as overconfidence, anchoring on recent news, weak base-rate reasoning, and inconsistent structured outputs. During development I plan to compare model combinations, evaluate performance on past resolved questions and MiniBench, and iteratively improve calibration and reliability based on observed errors.

**resilience-bot**

> The bot builds on the official template and adds three things. First, a multi-provider ensemble: Anthropic, Google and OpenAI models are called in rotation, two samples each, so that no single model's idiosyncrasies dominate. Second, an explicit outside view: before forecasting, each model must name a reference class of comparable past situations and its base rate, then adjust for the specifics of the question and for the status quo. Third, aggregation and calibration: the samples are combined by averaging log-odds (the geometric mean of odds) rather than taking a median of probabilities, clipped to \[2%, 98%], then passed through a logistic recalibration σ(a·logit + b) whose parameters are learned only from the bot's own resolved questions, with a prior pulling towards the identity and a cross-validation gate. The goal is to test whether a small, honestly validated calibration layer adds value on top of frontier-model ensembles. Background: I spent a year building and auditing a forecasting model for pari-mutuel horse racing (32,970 races, strict time-split validation). Most of the apparent edges turned out to be data leaks, and catching them taught me to trust only out-of-sample measurements — which is the spirit I want to bring here.

**theomachado-bot**

> The bot that I have called AXE is a library, not a prompt. Most of the work is carried out beforehand to the point where a question is even asked: various feeds and public APIs continuously fill a local store with news, official statistics stored as dated vintages, release calendars, prediction-market prices, court and legislative dockets, and open-access papers, all of which are indexed for full-text and semantic search and are linked to one another through shared entities, series, and dates. Upon a question being posed, the bot first breaks it down into a keyword plan and checks how similar previous questions had been broken down, then asks the library what it already has before carrying out any search on the live web. Live research (which includes web searches and news APIs) is only carried out on the areas that the library identifies as gaps, so that in the same ninety minutes more can be achieved and the evidence trail remains reproducible. Each forecast reads from a sealed evidence packet that includes an as-of timestamp and content hashes, and the decision record keeps track of the cruxes, the evidence used, and the prompt versions.

**Trigpoint**

> Trigpoint tests one idea: whether a group of forecasting models does better when each member sees different evidence instead of the same research summary. When several models read one shared briefing, their mistakes tend to line up, so averaging them helps less than expected. Trigpoint splits the evidence instead. One analyst reads only recent news. One sees only price data, prediction markets and a simple statistical model. One reads only the sources the question will be resolved from. One gets no evidence at all and reasons from base rates. They run on different models: Claude Opus 5, GPT-5.6 Sol and Gemini 3.6 Flash. Each question is first sorted by type, such as price thresholds, sports, weather, official counts or events. The type decides what evidence gets gathered and which two analysts go first. The other two are added only when the first two disagree, one fails, or a rule-based check catches an analyst citing something it wasn't shown or quoting a wrong price. For price questions, a statistical model estimates the chance of crossing the threshold within the question's date window. Forecasts are averaged in log-odds, pulled partway toward that model when one exists, and kept between 2% and 98%. Changes are tested only on questions that have already resolved. They're replayed with the clock set back to the original date and only the evidence that existed then. Repeating the same run several times shows how much scores move by chance, and a change is kept only if it beats that. Features are switched off one at a time to see which ones actually help. Results will also be compared against Metaculus's template bot running the same lead model.

**vezo**

> A bot that functions less as a pure probability engine and more as a living intelligence workbench. For each question, it ingests and synthesizes fragmented public evidence (news, data releases, expert commentary, related forecasts) into coherent and evolving mental models of the underlying situation. Instead of simply emitting a point forecast, the bot surfaces the most consequential uncertainties, maps competing causal interpretations, and shows how different assumptions shift the probability distribution, explicitly highlighting where the evidence is thin or contested.

**Cassandra-bot**

> CassandraBot is a six-model ensemble that has been running since the Spring season. Each question goes to Lightning Rod's foresight-v3 alongside five OpenRouter models, with prompts cycled across three reasoning perspectives: outside view (base rates and status quo), inside view (causal mechanisms and current trajectory), and devil's advocate (arguing against the consensus). Binary forecasts are aggregated by trimmed log-odds mean, multiple-choice by per-option median, then extremized in logit space.
> The extremization is horizon-scaled rather than fixed. A single factor applied to every question over-sharpens long-horizon forecasts where the uncertainty is genuine rather than an artifact of ensemble averaging, so the factor decays from 1.4 at short horizons toward 1.0 past a year.

## Summer 2026

**lakshya-bot**

> I am going to take an approach similar to the one highlighted in psychology of Intelligence analysis by Richard Heur Jr.  Essentially for all multiple choice questions, I will force the model to come up with a few hindsight explanations assuming something happened. For example question: When will US Iran was end: A- August-Dec 2026, B-After 2027 Jan C) before August
> I will generate an explanation of war ending in each. Then map out the chances of each of these events happening to get to the answer.

**Killian**

> Ora is a six-stage pipeline built on the thesis that scaffolding quality now matters more than model choice. Stage one is "resolution forensics": before any research, the bot adversarially parses the resolution criteria and fine print, fetches the content of every URL referenced in them, enumerates alternative resolution interpretations, and checks whether the outcome is already effectively determined. Stage two is a breadth-first research fan-out — news search (latest and archive strategies), provider-native web search on two different model families, and semantic search — followed by a gap-check pass that issues follow-up queries for whatever would most change the forecast. Every fact in the dossier carries a source date.
> Stages three through five: eight independent scratchpad forecasts across three model families, split between outside-view-first and inside-view-first framings, each required to compute an explicit numeric base rate. If the ensemble's spread exceeds a per-question-type threshold, the bot extracts the factual crux of the disagreement, runs targeted searches to settle it, and a supervisor model may override the statistical aggregate only at self-rated high confidence — reflecting the evidence that naive judge-style aggregation underperforms the simple mean. Final calibration: log-odds extremization at sqrt(3) with clipping to \[0.02, 0.98] on binary questions; per-option mean with flooring and renormalization on multiple choice; and for numeric questions, per-sample monotone-cubic interpolation in scaled space, a pointwise-mean mixture of the sample CDFs (widening tails rather than averaging quantiles), a small uniform blend as tail insurance, and strict pre-submission validation against the platform's CDF constraints.
> The deliberately less-common design choices: criteria-link forensics as a first-class pipeline stage, disagreement-triggered crux research before aggregation rather than generic re-sampling, and mixture-of-CDFs numeric handling. The bot is fully autonomous on a 15-minute polling loop with a complete audit trail of every dossier, reasoning sample, and submission.

**Oraky**

> Oraky is already live: it has forecast autonomously on a 15-minute tick since 2 July 2026 across the Summer 2026 FutureEval tournament and MiniBench, with about ten forecasts on the public record so far - a running bot you can check on Metaculus, not a proposal. What separates it from a typical prompt-an-LLM entry is that verifiability is engineered into the method, not promised: anti-look-ahead enforced in code, a research-versus-blind edge measured per question rather than assumed, and a forecast record stamped with the exact method version behind every number.
> The method runs two tracks with an explicit reconcile step. The BLIND track produces an evidence-first estimate with Claude Sonnet 4.6 - the prompt forces dated facts and counter-arguments before any number - and is given no market, betting, or crowd probability, so it structurally cannot anchor on the answer. The RESEARCH track runs the same estimator over a brief of fresh, dated web evidence; bookmaker and prediction-market domains (Polymarket, Kalshi, PredictIt, Manifold, Metaculus, oddschecker, and others) are hard-blocked, the prompt forbids querying odds, and the brief is guard-checked so any leaked implied price triggers a loud failure instead of silently poisoning the estimate. A RECONCILE stage weights the two tracks into the submitted probability while storing BOTH raw estimates (blind\_p and research\_p) on every two-track forecast, so the marginal value of research is measured question by question rather than assumed. Every forecast is written to an append-only log stamped with its method version, the project's core record is a separate hash-chained, externally-anchored ledger for tamper-evidence, and on resolution the bot auto-scores Brier, skill versus a coin-flip baseline, and the research-minus-blind delta-Brier - all under fail-closed daily and monthly budget caps in code that stop a run rather than degrade it silently.
> The two-track research layer went live 6 July; its first two research-backed forecasts already store both probabilities and show research materially moving the estimate (q44411 blind 0.62 -> research 0.38; q44378 blind 0.28 -> research 0.45). Rather than re-run a common design, the bot deliberately explores the enforced-anti-look-ahead + measured-research-vs-blind direction. The first tournament resolutions - and with them the first real Brier and research-versus-blind numbers - land in the coming weeks, and the resulting record, including a public paired blind/research calibration dataset, is offered back to the community.

**germanr-bot**

> germanr-bot is a cross-provider ensemble forecaster built around a calibration-first evaluation harness. Each question passes through triage (structured resolution-criteria parsing, an ambiguity score, and a shallow/deep compute dial), then research (typed queries over AskNews/Exa behind a code-enforced publication-date mask), then 5-6 reasoning members with distinct perspectives (reference-class, mechanistic, status-quo, market-aware, domain specialist, tail-risk) drawn from Anthropic, OpenAI, and Google models. Members pool via a Winsorized geometric mean of odds; high-dispersion questions trigger an adversarial pass in which one referee argues the aggregate is too high, one too low, and a supervisor reconciles, rejecting claims not traceable to cited sources.

**xiejohnm-bot**

> 1\. Forecasting Philosophy: My bot focuses on the post-processing side of forecasting rather than novel research scaffolding: it treats the LLM as a noisy evidence-to-probability converter and puts the engineering effort into calibration machinery around it. Raw estimates come from a small ensemble of Claude samples (median-aggregated) fed with reference-class base rates and news evidence; the distinctive part is what happens after. Every raw probability passes through an explicit calibration correction and an evidence-strength gate: when the gathered evidence is thin, the forecast is deliberately shrunk toward the base rate or market anchor, because log scoring punishes confident tails far more than it rewards them.
> 2\. Research Background: The design comes from earlier private research on real prediction markets (weather-settlement markets), where we measured systematic calibration biases such as the favorite-longshot effect and found that disciplined abstention-style shrinkage beat aggressive point estimates over months of walk-forward testing.
> 3\. Calibration & Validation: The bot ports that playbook to Metaculus questions: market prices and community-visible forecasts are used as anchors where the rules allow, corrected by their measured bias curves rather than copied. Per-domain calibration is tracked across MiniBench editions, and improvements are only shipped when they win on disjoint validation slices, not on a single leaderboard bump.

**noema-bot**

> The bot combines a multi-persona reasoning ensemble with an explicit outside-view step. For each question it first estimates a base rate from reference-class reasoning, then runs several distinct forecasting personas with different priors and risk stances, and aggregates them with shrinkage toward the base rate weighted by evidence quality. Research context is gathered automatically before reasoning.
> Rather than relying on a single model call, the system treats forecasting strategies as competing hypotheses. It keeps an attribution ledger tying each forecast to the strategy that produced it, and once questions resolve it scores strategies on out-of-sample resolved questions using proper scoring rules, then reallocates weight toward whichever is calibrating better over time.
> The whole loop is autonomous and scheduled: it forecasts new questions on a recurring basis, handles every question type defensively so a single malformed input can never end a run, and improves itself between rounds based only on resolved-question performance

**aroswift-bot**

> Most bots this season and last seem to follow the same basic recipe: run several forecasts, combine the results, adjust confidence, and separate inside-view from outside-view reasoning. Rather than build another version of that, I'm focusing on where the published benchmarks suggest bots lose the most points: multiple-choice questions.
> The bot-vs-pro gap appears to be much larger on multiple-choice questions than on binary ones, yet few bots seem to target that directly. My first pipeline draft reduces option-order bias by forecasting each question with several shuffled answer orders and averaging the results. It also forces probability onto long-shot options and normalizes the final probabilities so they always sum to 100%.
> Calibration is empirical rather than guessed. I use previously resolved questions as a holdout set, then fit a per-bin recalibration map in python. For example, if 33% forecasts only resolve true around 20% of the time, the system learns to shift similar forecasts downward. An objectivity check prevents mechanical questions, like dice odds, from being incorrectly pulled toward historical base rates.
> For binary and numeric questions, I'm maintaining a proven ensemble-and-aggregate approach, with code execution for quantitative steps. Also, I'm entering both MiniBench and the primary tournament, using MiniBench’s two-week cycles to tune the recalibration and multiple-choice debiasing.

**praescientia**

> We plan on using OpenEvolve's LLM Evolutionary Search to evolve bots for the forecasting competitions (https://github.com/algorithmicsuperintelligence/openevolve). We are interested in applying this approach to forecasting by evolving bot architectures, prompting strategies, research pipelines, and calibration heuristics against historical Metaculus questions before deploying fully automated entrants.

**Sloth**

> Sloth is the forecasting arm of SlothAI, a personal autonomous AI system that already routes work across multiple frontier and open models. For each question the bot retrieves fresh context (AskNews and Exa search), fans the question out to several frontier models with multiple samples per model and several distinct reasoning angles, then aggregates the samples into a single calibrated probability with an auto-generated reasoning comment.
> The differentiating bet is infrastructure: the bot lives inside a self-improving agent loop with a persistent scoring ledger. Every submission is recorded, scored on resolution, and fed back into prompt/aggregation selection, so the ensemble configuration is chosen by measured Brier improvement on the community feed and MiniBench rather than by intuition. The system already does this end-to-end for the Numerai tournament (40+ scored submissions) and the Metaculus community feed (43 submissions).
> Plan for the season: harden the retrieval + multi-sample aggregation pipeline, iterate bi-weekly on MiniBench, and submit to the seasonal tournament only configurations that beat our community-feed baseline. We will publish a postmortem on what the scoring ledger says actually moved calibration.

**Blueberry Pi**

> Blueberry Pi is a Summer 2026 FutureEval forecasting bot built around a stronger Deep Research pipeline. We previously placed 16th in the Metaculus 2025 Fall bot tournament (OliveTestBot) with a multi-agent debate forecasting stack, and the main lesson was that upstream research quality often dominates final forecast quality: decomposing the question correctly, identifying decision-relevant variables, resolving conflicting evidence, and updating when new information arrives.
> The new system centers on EvoGraph, a persistent evidence graph/wiki of entities, mechanisms, sources, contradictions, and open questions. EvoGraph decomposes forecasts into decision-relevant components, searches for missing evidence, updates beliefs when stronger evidence arrives, and routes later research toward weak nodes, single-source claims, disconnected clusters, and unresolved contradictions. For example, an AI supply-chain forecast would decompose an NVIDIA Rubin rack into GPUs, HBM, PCBs, liquid cooling, power modules, optics, assembly capacity, and supplier constraints, then identify likely bottlenecks and resolution-relevant signals.
> We will combine this graph-based research memory with our earlier multi-agent debate method, a MiroFlow/MirrorFlow-style research implementation, and a lightweight forecasting-skills library distilled from forecasting and intelligence-analysis practice: source triage, base-rate search, timeline reconstruction, market/expert expectation search, causal-chain analysis, component-level decomposition, adversarial red-team review, and calibrated final forecast aggregation. The output remains a standard Metaculus forecast and reasoning comment, but the pipeline should produce deeper evidence collection and more transparent uncertainty tracking.

**dtfr-bot**

> dtfr-bot is the forecasting engine of a non-commercial OSINT analysis demo. The demo already implements a "centaur" workflow: the AI proposes a probability with sourced FOR/AGAINST reasoning, drivers and indicators; the human analyst adjusts and decides; both tracks are Brier-scored separately. FutureEval will serve as the external, independent evaluation harness for that engine.
> Architecture: retrieval-augmented reasoning. For general questions, AskNews/web retrieval. For China-related questions, the bot additionally queries a curated corpus of \~13,500 documents from vetted Chinese-language and international sources. The scaffold decomposes the question, establishes base rates, weighs recent evidence for/against, and maps an ICD-203-style verbal band to a calibrated probability; 2-3 model runs are aggregated (trimmed mean).
> The experiment we care most about: submitting a main corpus-augmented bot and one secondary baseline bot (identical scaffold, no domain corpus) to measure whether curated regional OSINT measurably improves forecast accuracy on China-related questions versus a generic pipeline. Findings feed the demo's recalibration layer, and we plan to publish a postmortem at season end.

**apemble-bot**

> I'm building a multi-agent forecasting pipeline for the Summer Bot Tournament. It will include a planner agent that reads the question and identifies what information is needed to resolve it, a research agent that retrieves relevant news and data, an ensemble of forecasting agents that each reason from different first-principles frameworks and argue against each other's estimates, and a final aggregator that synthesizes the outputs into a calibrated probability.

**pancratic-bot**

> pancratic-bot is a calibration-first forecasting bot. Rather than relying on a single large model's gut probability, it runs each question through structured base-rate research and a three-lens reasoning ensemble: an outside view (reference-class / base-rate framing), an inside view (causal, question-specific mechanics), and a status-quo/inertia lens (what happens if nothing changes). Each lens produces an independent probability and rationale.
> The lens outputs are combined by a consensus-aware aggregation layer operating in log-odds space — taking the median and applying agreement-scaled extremization: when the lenses independently converge, the bot grows more confident; when they diverge, it stays humble and pulls toward uncertainty. The same machinery handles binary, multiple-choice, and numeric (percentile/CDF) questions. The edge is disciplined evidence aggregation and calibrated updating, not raw model size.
> The thesis: most bots lose points to overconfidence and poor calibration, not lack of knowledge. By forcing diverse reasoning paths and only extremizing when they genuinely agree, pancratic-bot aims to be better calibrated than any single-prompt baseline while keeping per-question token cost low enough to cover the full tournament on sponsored credits.

**GreeneiBot2 (top of leaderboard in Spring)**

> An async Python bot that automatically forecasts Metaculus questions. It fetches open questions from configured tournaments, then runs parallel research before forecasting: AskNews (news context and deep research), Perplexity (sonar-reasoning-pro and sonar-deep-research), GPT-5 search via OpenRouter, and link assistants that screenshot resolution links and process them with GPT-5 or BrowserUse. The bot uses LLMs (OpenRouter/OpenAI via litellm) to produce forecasts for binary, multiple-choice, numeric, and discrete questions, with structured CDFs (percentiles and spikes) for numeric/discrete, validation against platform rules, and multi-run aggregation.

**henryb-bot**

> At a high level, I want to have one model summarize the prompt and make a plan for figuring it out, another do research, a few forecast and argue against each other, and one more compile everything and make a decision. I will test which models are best at each based on how close they get to the community prediction on test questions. After I collect data I will finalize the bot, with each model doing its specialty/ies, and watch it compete in the Summer AI Forecasting Benchmark Tournament.

**MASXAI-BOT**

> MASX-AI is a multi agent forecasting pipeline organized around a Council of Doctrines. Each question is first parsed into a structured hotspot context (geopolitical actors, domain pillar, base rate signals from Polymarket and web search). A router agent selects the doctrines most relevant to the question from a curated registry covering international relations theory, conflict dynamics, economic statecraft, and other domains. Selected doctrines become parallel council agents, each reasoning through its lens against retrieved doctrinal text and live news evidence. A forecaster aggregates the council outputs via log odds, applies signal driven recalibration (calibration version, close approach, search drift), and clamps the final probability to a Brier-safe range before submission.
> Evaluation and iteration are first class. Every submission persists raw post snapshots, structured reasoning records, and per agent doctrinal analyses to Supabase. This lets me detect question text drift across versions, compare my forecast to the community prediction at close, and run targeted reforecasts when a calibration patch ships without re-running questions whose state has not changed. MiniBench is the fast feedback loop (every 2 weeks, \~60 questions) and the seasonal tournament is the slow ground truth. Bot reasoning is posted as a public comment per Metaculus tournament transparency rules, with a minimum content threshold to avoid posting boilerplate.
> The novel piece is the doctrinal council. Rather than a single chain-of-thought or a generic ensemble of identical prompts, each council agent is anchored to a specific theoretical lens (e.g. balance of power, prospect theory, hegemonic stability) and is forced to argue from that frame. Aggregating across heterogeneous frames is intended to produce better calibration on geopolitical questions than monoculture ensembles, while the snapshot store lets me measure whether the doctrinal diversity actually pays off versus a simpler baseline. Credit budget primarily funds the council fan out (Anthropic Sonnet and Google Gemini for council members, frontier model for final aggregation), with the search and embedding paths running on local models or independent search providers.

**Calibra**

> My bot is a research-oriented Metaculus forecasting system designed to evaluate how large language models perform on live forecasting questions when supported by structured prompting, evidence retrieval, aggregation, and calibration. The current working version uses the official Metaculus forecasting-tools/template infrastructure and a Superforecaster-style prompt adapted from prior forecasting-prompt literature. It handles binary, multiple-choice, numeric, and date questions, and logs each individual model sample, parsed forecast, aggregation output, prompt version, model used, and research mode.
> The first baseline uses a no-external-retrieval setup, where the model produces a research-style summary from its own parametric knowledge before forecasting. The next stage is to add external evidence retrieval, likely through AskNews or a similar provider, so I can compare no-external-retrieval forecasts against evidence-grounded forecasts in a controlled way. I also plan to test aggregation across multiple samples and later across multiple models, while monitoring calibration and overconfidence.
> The project is connected to my PhD research on LLM forecasting, evidence access, calibration, and prediction-market-style evaluation. My goal is both to compete in the tournament and to generate useful research data about when retrieval and calibration improve AI forecasting performance.

**signal2noise**

> Signal 2 Noise is a Python forecasting bot competing in the Metaculus AI Benchmarking Tournament. Rather than asking a single LLM "what's the probability?", it runs a structured multi-stage pipeline designed to anchor estimates on empirical evidence before any LLM reasoning begins. First, 13 reference-class modules inject historical base rates (electoral incumbency, legislation passage rates, armed conflict frequency, clinical trial outcomes, etc.) derived from structured databases. The LLM then reasons from that empirical anchor rather than from prior beliefs alone.
> Evidence gathering runs in parallel across 40+ live data sources — prediction markets (Polymarket, Kalshi, Manifold, Betfair), news search (AskNews, Exa, Perplexity, Serper), economic data (FRED, IMF, OECD), and domain-specific APIs (ACLED, USGS, FDA, SEC EDGAR, CourtListener). A ReAct-style gap-filling loop identifies missing evidence and queries targeted follow-up searches. The final probability is a logit-weighted blend of a base-rate anchor (P1), a current-signals pass (P2), and a resolution-criteria adjustment (P3), with calibration multipliers applied per domain based on tracked historical performance.
> The bot runs hourly on a Hetzner server, forecasts up to 12 questions per run, and publishes its full forecast report and a weekly blog at signal2noise.uk. The codebase is open-source at github.com/signal2noise2/metaculus-bot.

**Panshul42**

> Agentic search (with live search tooling) with prompt guidance about what to look for, consolidation, parallel one-shot forecasting, ensembling, summarization.
> Thinking of making forecasting multi-shot potentially or allowing forecasters to re-call the search system in some ways if it wants to find more information.
> (This sounds simpler than my Q2 bot from last year, and it is, because much of the complexity needed for google search/scraping can be easily and reliably abstracted to native LLM provider web search)

**silver-surfer**

> For now, just the basic framework. First thing I'd change is the search. I'd likely connect a few data toolkits (like pred. market data, websearch, etc.) and my own research loop. Second improvement would be a best-of-N multi-model predictive engine. Basically reduce impact of hallucinations and aggregate model variation. After that, I'd likely implement concurrent research agents (for greater breadth) and a scratch-pad so agents can make a more in-depth reports prior to prediction. I honestly have a ton of ideas but have no sense of what the real bottleneck is yet.

**dazzer\_aib\_tournament**

> Bot works like an agent harness, starts with a simple memory.md and backtested, then reflects and improves its own instructions and methodology. Interactive graphs are used to represent the outcome of searches to save context length.

**tejvn-bot**

> My bot is a fork of NoStream's open-source forecasting bot, which I intend to extend with additional research providers, and an improved aggregation pipeline. The core architecture aggregates predictions from frontier reasoning models across OpenAI and Anthropic, using conditional stacking to resolve disagreements: when base forecasters diverge significantly, a meta-model extracts the crux of disagreement, runs targeted research to resolve it, and synthesizes a final prediction, a design intended to capture complementary strengths of different model families while actively handling the cases where they conflict.
> My current focus is grounding quantitative reasoning in agentic computation. Currently, the pipeline reasons about numbers in prose, which introduces systematic errors in arithmetic, distribution fitting, and base rate calculation. I plan to integrate agentic code execution so the forecasting agent can write and run Python to ground quantitative claims: (EDA, regression, simulation, etc)
> On the data side, I'm incorporating other structured sources (IMF/World Bank, BLS, Census Bureau, Tradier), and I plan to add Polymarket and Kalshi as research inputs with the goal of weighting prediction markets by their relevance to each question's resolution criteria and by liquidity, volume, and spread.

**smingers-bot**

> My current bot is a ReAct loop for research with an ensemble receiving its output. I am interested in pursuing more discrete research tracks for decomposed sub-questions, code execution for forecasters to improve quantitative reasoning, prompting and other strategies to produce greater meaningful divergence in the forecasting ensemble (that remains grounded in the fact pattern).

**Beskar-Bot**

> It is a multi-agent platform designed to analyze prediction markets by simulating a council of experts. At its core, the system uses specialized AI analysts that each supported by their own sub-researcher agents to independently gather data and form verdicts on a specific prediction question. This decentralized research approach ensures that multiple perspectives and domain expertise are considered before any conclusions are drawn.
> Once the independent verdicts are in, a "leader" agent synthesizes the findings into a preliminary recommendation. To ensure rigor and prevent bias, this recommendation is immediately challenged by a dedicated Adversarial Reviewer (modeled after the CIA's Red Cell protocol). The leader must explicitly address every critique raised by the adversarial reviewer before issuing a final, calibrated probability and verdict.
> Finally, the platform is built for continuous learning and dynamic updates. An optional "re-analysis loop" can periodically check for new information, using an "Evolution Gate" agent to filter out duplicate data and only trigger updates when genuinely material evidence emerges. When a prediction market finally resolves, a built-in calibration system grades the AI's performance to adjust source credibility and improve the system's accuracy over time.

## Spring 2026

**GreeneiBot2 (Spring 2026)**

> GreeneiBot2 — An async Python bot that automatically forecasts Metaculus questions. It pulls open questions from configured tournaments (e.g. spring-aib-2026, minibench), then runs parallel research before forecasting: AskNews for recent news and deep research, Perplexity (sonar-reasoning-pro and sonar-deep-research), GPT‑5 search via OpenRouter, and link assistants that screenshot resolution links and process them with GPT‑5 or BrowserUse. The bot uses LLMs (OpenRouter/OpenAI via litellm) to produce forecasts for binary, multiple-choice, numeric, and discrete questions, with structured CDFs (percentiles and spikes) for numeric/discrete types, validation against platform rules, and aggregation over multiple runs.

**caliperbot (Spring 2026)**

> I'm calibrating an LLM's predictions by making many queries on the distribution, or conditional distribution, of random variables related to the target question, and solving for the distribution over all variables which best fits the answers, under some prior on the trustworthiness of a typical LLM prediction. Right now that means fitting a maximum entropy model with weighted constraints, but I'll likely switch to a Monte Carlo approach to get more control over the prior.

**geemus+bot (Spring 2026)**

> Currently it compiles initial research, fans out to multiple forecasts, does a single round of delphi, and then a single consensus is built from the delphi-updated forecasts. I'll probably stick more or less to that structure, but add additional models/providers and also explore more agentic/tool-usage based possibilities.

**nostreambot (Spring 2026)**

> [similar to present](https://github.com/No-Stream/metaculus-bot): mostly using the very classic technique of ensembling. i initially wanted to do some more involved stacking methods but decided against it due to variance/risk, so just sticking with median agg mostly likely. i am thinking about building some more agentic research capabilities this quarter, which will obviously increase cost and complexity but should be fun and just maybe useful!

**r2d2 (Spring 2026)**

> r2d2 implements recursive Fermi decomposition. Rather than feeding the full question to a single LLM, r2d2 first breaks questions into 3-5 tractable sub-questions, researches each independently, and synthesizes estimates bottom-up. When a sub-question is still too complex, it recurses and decomposes further, stopping when questions become tractable (estimable with reasonable confidence from research + base rates). The architecture explicitly implements all 9 steps from forecasting methodology literature (Tetlock, Mann): base rate anchoring, risk vs uncertainty classification, logical bounds reasoning, and adversarial premortems. Research pulls from multiple sources (AskNews, Perplexity, prediction markets) to reduce single-source bias. Final forecasts are ensembled across 3-5 runs with varied prompts (outside view vs inside view emphasis) and aggregated via median. The hypothesis is that structured decomposition + adversarial challenge can close the gap between bots and pro forecasters, since this is what pros do mentally but bots typically skip.

**hugo0-bot (Spring 2026)**

> Pythia is a multi-pass ensemble forecasting bot built on Claude Sonnet 4.5 for speed and Claude Opus 4.6 for deep analysis, as well as other models during ensembling. For each question, it runs an independent research phase followed by multiple parallel prediction runs with temperature-based diversity. Predictions are aggregated via median and calibrated through logit-space extremization (factor 1.3-1.5), a technique shown to be the highest-ROI single intervention for LLM forecasting calibration. The architecture is inspired by the AIA Forecaster (Bridgewater/Metaculus SOTA, Brier 0.108) and the Q2 2025 Metaculus tournament winner. Key design choices include blind scoring (no anchoring to community predictions during reasoning), inside/outside view decomposition (base rates vs. current evidence), and multi-run ensembling to reduce variance. The bot handles binary, numeric (201-point CDF via percentile interpolation), and multiple choice questions. Pythia was originally built as a Polymarket trading system and is being adapted for Metaculus competitions. The core forecasting logic transfers directly — the main difference is structured resolution criteria (which actually makes Metaculus questions easier to reason about). We plan to add web search integration (AskNews/Perplexity) for real-time information retrieval and iteratively improve prompts based on MiniBench feedback loops.

**cognos-bot (Spring 2026)**

> My bot is built on CognOS, a distributed cognitive operating system that I developed.

**amigon21-bot (Summer/Spring 2026)**

> MiniSim is a multi-agent swarm prediction engine with variance-based routing. For each question, it generates 10-20 diverse agent archetypes with temperature stratification. It measures initial agent variance, then routes: low variance → single frontier model call (agents already agree, deliberation adds noise); high variance → full 4-round structured deliberation (evidence exchange, critique & rebuttal, final forecast). Predictions are aggregated via calibrated confidence-weighted extremization (alpha\=1.5), with Platt scaling calibration applied at inference. So far, all development has been done on local Ollama (Qwen 2.5 14B on Apple M4 Pro, \$0 API cost). Over \~40 hours of on-device inference across 1,000+ LLM calls, we've validated: the router correctly dispatches 56% of questions to single-model mode and 34% to full swarm; extremized swarm achieves Brier 0.017 on a 10-question benchmark (53% better than single LLM); offline calibration on 544 questions shows ECE 0.059 (beating market ECE of 0.073). We've already submitted 50 live forecasts to the Spring 2026 tournament using Qwen 14B. Requesting credits to test with Claude Sonnet and o3 — Metaculus research shows SOTA model choice is the single biggest accuracy differentiator, and we want to measure whether frontier models justify the cost over our free local setup. The system is fully open source with 70 tests, a FastAPI production API, and integrations with 5 prediction market platforms.

**emdash\_Delphi (Spring 2026)**

> The bot uses a fairly straightforward aggregation approach to combine multiple searches, models, and prompts into a single forecast in a scalable way. The structure has two major components. The first component is a multi-stage research + forecasting pipeline that runs a single web search using a lightweight model (currently GPT-5-mini, with agentic search), which is fed into a researcher that adds context using a frontier model (GPT-5.2), and then used to make N predictions using frontier models. The second component combines both of these steps into a model call of a frontier model that includes both research (with agentic web search) and the forecasting steps. These forecasts are combined using means across all forecaster modules, which currently use a flat weight but planning to upweight the end-to-end models as they are more expensive on a per-run basis. The focus of the last tournament was on reducing hallucination rate, which I feel was modestly successful but want to continue to iterate on. The next planned addition is to test the incorporation of a judge LLM which evaluates the plausibility of the above steps to identify potential hallucinations or disagreements and dynamically adjust the weights.

**jonah-ramponi-ai (Spring 2026)**

> I don't think I have any grand plan for now. I think some things which make sense are (A) breaking down a problem into subparts, and researching those independently (probably something like sub-agents) (B) trying to get away from the fact that LLMs are bad with numbers (I plan to use logprobs from the api, rather than asking the llm to spit out a number. Maybe I'll train a simple model for "gamesmanship" to try better tune the confidence. (C) Certainly web searching I imagine once I've spent a week on it, I'll totally rethink all of the above, and my approach will be nothing like I think it should be now! Not quite sure how I will achieve it yet, but I think an ensemble approach typically makes a lot of sense for forecasting problems. I don't think I'll go down the route of having lots of different agents who debate things.

**hanshalg (Spring 2026)**

> The plan is to have wide array of information sources (Exa, Google, Perplexity, AskNews, etc.) for research whilst aggressively pruning 'low-signal' pieces to combat context rot. In an 'agentic' loop, the bot will research and generate an outside and inside view using various prompting ideas. I will experiment with adding extra tools such as running code, reading pdfs, etc. This outside and inside view will be fed into a frontier model and sampled numerous times with the final forecast being the median.

**joy.void.joy-bot (Spring 2026)**

> I'm using Claude Code agent sdk as a thinking loop [https://platform.claude.com/docs/en/agent-sdk/](https://platform.claude.com/docs/en/agent-sdk/) . It has access to a lot of thinking tokens, many different tools (builtin websearch/webfetches, notes of past forecast reviews, google trends, yfinance, metaculus/manifold exploration etc). I am using Claude Code to iterate on this project, review past forecasts, add more tools, refine prompt/refine the ontology of different subagent, etc... at different levels of recursion. The project is moving fast and I'm working on it daily.

**Pirohuni (Spring 2026)**

> Expanding on the idea of adding personality diversity, LLM diversity and other variations: I plan on expanding the diversity of responses given by different LLMs and have a final LLM aggregate the opinions provided while giving weightage to each one to get the best result.

**RRRRRR-bot (Spring 2026)**

> My ideal bot is designed to avoid two common problems in AI forecasting: overconfidence and overreaction. Instead of jumping straight to a final probability, it first asks “what usually happens in situations like this?” and sets a reasonable baseline. Only after that does it look at recent news or signals and adjust the forecast, with limits on how big the adjustment can be. The bot also stress-tests its own prediction by asking a simple question: “what would need to go wrong for this forecast to be wrong?” If the outcome depends on a single fragile assumption, the bot becomes more cautious. If it would take several unlikely things failing at once, the bot allows higher confidence. This helps keep probabilities realistic rather than extreme. Overall, the goal is not to sound smart on any single question, but to stay well-calibrated over time. The bot is built to perform well under log scoring and time-averaged evaluation by staying stable early, updating when evidence truly changes, and avoiding rare but costly prediction mistakes.

**murali11-bot (Spring 2026)**

> As this is my initial trial I want to understand the base capabilities. I've created a debate simulator in the past using Claude and would like to do something similiar here for forecasting. Part of this would also be trying to breakdown the topic into smaller sections and then build up a picture from there.

**djuarez-bot (Spring 2026)**

> The bot will use AskNews and Perplexity to research the topic where AskNews focuses on finding more recent sources and Perplexity focuses on more historic data. This is aggregated and passed to the forecasters which will be 2 models (not too sure which ones, but likely a Anthropic model and an OpenAI model). One of the models will focus on being objective by using numbers to back up its statement, while the other model's prompt will ask it to focus on human sentiment on what the answer could be. These will be aggregated (normalized) and submitted.

**Chineme (Spring 2026)**

> The bot automatically forecasts Metaculus questions by first gathering web research (via Tavily) and summarizing the key evidence with an LLM. It then prompts an LLM to reason through scenarios and produce structured probability forecasts tailored to the question type (binary, numeric, multiple choice, etc.) slightly extremizes binary probabilities away from 50%

**VIGILJANUS  (Spring 2026)**

> A calibration-focused forecasting bot built on Claude Sonnet 4.6 (via OpenRouter), designed to prioritize epistemic accuracy over speed. The bot uses a structured research-then-forecast pipeline: for each question, it first generates a research report analyzing key factors for/against, base rates, and status quo assessment, then produces multiple independent probability estimates that are aggregated into a final forecast. The core design principle is "good forecasters put extra weight on the status quo" — the bot is explicitly prompted to resist novelty bias and anchor on how slowly the world typically changes. Development follows an iterative approach using MiniBench as a 2-week feedback loop. Each cycle: review calibration curves, identify systematic biases (overconfidence, domain-specific blind spots), adjust prompts and research strategies, re-run. Manual review of outputs in the first 2 weeks to build intuition for failure modes before shifting to automated 30-minute forecast cycles. The bot logs every forecast with a reasoning trace (goal, options considered, decision rationale, counterfactual) to enable systematic post-mortem analysis rather than ad-hoc prompt tweaking. This bot is the first module of a larger private forecasting engine (SYGNITO) focused on technology, AI, and crypto domains. Long-term goal: evolve from a standalone script into a dedicated forecasting agent within a multi-agent orchestration system, where research, calibration review, and forecast submission are handled by specialized sub-agents. The tournament serves as both a proving ground for the forecasting methodology and a source of real-world calibration data.

**Big-A-Bot (Spring 2026)**

> I want to build a full agent taking advantage of modern features like mcp, skills, and more. While I have some experience building AI agents this is my first time building a forecasting bot, so I plan on doing a lot of tinkering with what scaffold works the best. In particular I want to try and build tools that offload parts the bot finds difficult about forecasting (maybe concrete probablistic modeling), and allow it to act as more of an orchestrator/high level thinker. Again, I'm not entirely sure what works - I will have a lot of trial and error. I have yet to fully review past approaches, so that will also impact the direction of the project.

**pshah-bot (Spring 2026)**

> Traditional prediction market models often assume homogeneous participant behavior or static event probabilities; however, in reality, beliefs and behaviors evolve dynamically. By modeling these shifts as latent clusters in participant trading and price data, the bot identifies distinct market “modes” that correspond to collective sentiment regimes — such as optimism, pessimism, or uncertainty — and adjusts its betting or position-taking strategies accordingly.

**mmBot (Spring 2026)**

> My bot first does research planning by decomposing the question and generating targeted search queries, scraping and summarizing the most relevant articles (llm-driven). It then classifies the question and routes it (also llm-driven) to one of several specialized context pipelines (eg. stock, sport, election, etc). Finally the optimized prompt and context is used to prompt an llm for a prediction several times which are then ensembled into the final forecast.

**CumulativeBot (Spring 2026)**

> My strategy has been scale: I have a lot of research methods, I do lots of pre-processing of research, and I use a very wide range of prediction methods. I built my own deep researcher with the atomic-agents library. I had mixed success: While cheaper on paper, pretty much every hallucination could be tracked back to it, it look a long time, failed a lot, and wasn't actually that much cheaper in practice. I'm planning to make a new one if I have time, or just move back to greater reliance on Perplexity, Exa, etc. Right now, for prediction, I have a system that uses generated perspectives for direct prediction + monte carlo sims, a system that uses tailored prompts based on the question (turned out to be quite expensive, so I turned that down), a deprecated agentic pipeline (expensive and underperformed; don't bother), and a system that takes base-rate-only research, generates a base rate, then updates it with contemporary information. I also experimented with "distilling" it into a Q-n-A prompt form and asking it, getting it to focus on the most vital things, but that was a bizarre idea I made in response to a specific failure, and I removed that after a while. I also try to veto obviously bad predictions with a set of sanity checks. This lead to false positives and a low participation rate last tournament - it takes a lot of tuning. I still think it's a useful thing to have, and I'm considering making it identify question irregularities (e.g. some of the questions being dated for the wrong year in the fine print) for the next tournament.

**PressPause (Spring 2026)**

> I would like to do experiments with Multiple Choice questions, because they seem to be a weakness for the bots. In particular, I want to run the same model on the same question but with different orderings on the replies to see how that affects the answers. I also want to see which difference it makes to ask the model to ensure the replies sum to 100% vs. ask it to not think about it at all (and then normalize in python) vs. first not think about it and then normalize. I also want to do calibration (in python rather than by the bot): e.g. if only 20% of questions that are estimated to 33% come true, then answers of 33% should be adjusted to 20%. However, on questions like "What is the probability that a dice throw ends in 5 or 6" I do not want to adjust 33% to 20%, so I would like to ask the bot how objective/precisely modeled the answer it, and adjust the calibration based on that.

**Quail (Spring 2026)**

> I'm taking an ensemble forecasting approach, with 1 parser, 5 researchers, 7 forecasters, 3 challengers and 1 summariser.

**q\_forc+bot (Spring 2026)**

> My prompt forecast strategy is multi-scenario for all question types, which is generally working well. Aggregation The forecasting-tools median approach for numerical CDF aggregation conflicts in a negative way with my bot generated scenarios. I plan to make an improved CDF aggregation function. I also intend to make a binary forecast aggregation function that naturally (slightly) extremizes the scenario forecasts. Diversification Secondarily I want to diversify models, aggregating forecasts from more than one model. I would also like to moderately augment AskNews results with additional data.

**Rexhaif (Fall 2025)**

> I want to focus on collaborative multi-agent system with a very relaxed hierarchy (a bunch of collaborating bots + single summary/answer agent). Each agent will be equiped with same tools but will use them differently according to it's role prompt - i.e it will plan analysis differently, focus on different areas, collaborate in it's own way. Role prompts will be generated as well. The rationale here is that simulating diverse perspectives will boost overall reliability and accuracy of the prediction.

## **Fall 2025**

**hb\_bot (Fall 2025)**

> I analyzed the results of the Q2 tournament to see, if some models may perform better at specific topics or question types. The bot would first categorize each question into a topic, e.g. "financial", "political", "political usa",... and depending on the topic a different model (and prompt) will be used. Currently I am thinking of using o3 for numeric and multiple-choice questions, o4-mini-high for binary questions. Political questions outside the usa will be forecasted by o3 and  financial binary questions related to the US will be forecasted on by o1high. I may adjust the "ratio" like doing 3/5 runs using o3 and 2/5 runs using o4.

**Upskillbot (Fall 2025)**

> Upskillbot is a specialized forecasting agent focused on questions related to the future of technology, education, and labor markets. Its primary goal is to not only produce accurate probabilistic forecasts but also to generate clear, educational rationales that explain the key drivers behind its predictions, aligning with our organization's mission to promote upskilling.
> ​The bot will employ a two-stage process. First, it will use a cost-effective model like GPT-4o-mini to act as a "Research Analyst," scanning and summarizing relevant news, research papers, and data sources for a given question. This summarized context is then passed to a more powerful "Reasoning Engine" powered by GPT-5, which synthesizes the information, evaluates different potential outcomes, identifies key uncertainties, and constructs a final forecast and detailed explanation.

**MarcosBot (Fall 2025)**

> My plan involves a multi-stage approach. Initially, I aim to assess each Metaculus question's complexity to guide a tiered research strategy. This assessment will involve generating a complexity score (1-10) for each question by averaging the outputs of o3 twice. Based on this score, research will be tiered: minimal AskNews searches for scores 1-2, standard AskNews searches for scores 3-6, and a combination of standard AskNews and AskNews DeepNews for scores 7-10.
>  
> Following the research, forecasts will be generated by o3 run three times and 4o Mini four times. I plan to use a custom prompt that encourages detailed probabilistic reasoning, partly by breaking down complex questions into simpler sub-questions for more accurate forecasts taking inspiration from Squiggle AI. I will then aggregate these seven predictions by averaging the middle five runs, giving double weight to the o3 model's predictions. I also plan to use GPT-4o Mini for summarization tasks, for generating comments on Metaculus.

**SynapseSeer (Fall 2025)**

> The bot will be similar to previous iterations, with a focus on prompts that help ensure the question is being evaluated appropriately and comprehensively. It includes both a base analysis LLM and an evaluator LLM to make a final forecast.

**NeoTartine (Fall 2025)**

> I'll replicate some idea I've read about on Arxiv, the main thing is to create a synthetic panel with a large-ish number of agents. Their answers and then averaged to create a final prediction.

**ferraijv (Fall 2025)**

> TBD; my initial thought is an ensemble where multiple bots give their recommendation to a final bot. The final bot reads the different arguments from each output and makes a final decision

**Raziel (Fall 2025)**

> This bot is going to be testing a series of specialized techniques for probabilistic prediction namely:
> \- meta modeling
> \- automated prompt tuning
> \- subproblem factorization
> Although the exact strategy will change based on what is effective :)

**botduke (Fall 2025)**

> The core of the bot's logic is a multi-agent AI system. One AI agent acts as a proponent, another as an opponent, and a final "synthesizer" agent evaluates their arguments and the source data to produce a final probability. This simulated debate is designed to reduce bias and create a more robust forecast before the bot decides whether to make its high-risk prediction.

**htimm\_bot (Fall 2025)**

> There is no reason that human forecasters should consistently be able to beat out LLMs forecasts, especially with modern reasoning models. To me this means either there is a lack of or misinformation being passed to LLMs or that some models are better fit at answering questions in specific fields.
>  
> The best way to counter this is to take the average across multiple LLM expert forecasters (passed with similar prompts), similar to the way the expert forecasters do. While this has been done before, the usual errors are that the exact same prompts are passed for some small n or multiple forecasts are given in one response. Additionally, only recent headlines are passed in most studies - considering most models have a cutoff much earlier than these headlines, they might be missing something critical, which would throw off predictions.
>  
> Since the expert forecasters all share a variety of backgrounds, it makes sense to replicate bots that are "experts" in their field but vary so the average converges to the true estimation.  I plan to create multiple agents in both finding recent relevant news and prior news then feeding it to multiple agents with different backgrounds but "experts" in their field.

**jonahsingerbot (Fall 2025)**

> I plan to systematically test changes to my 9th place bot from last tournament (GPT o3 with Ask news and Deep news) to see what configuration yields the best results. Below I have an order in which I plan to test using the provided community benchmark tool to test.
>  
> Reasoning model testing:
>  
> Test different reasoning models individually GPT (GPT-5 reasoning, o3), Gemini 2.5 pro, Claude Opus 4.1.
> I also would like to try GPT 5 pro if it is possible with in the budget.
>  
> News retrieval:
>  
> Ask News + Deep news (I used last tournament)
> GPT search api.
> Exa.ai
> Perplexity
> Test searching directly for questions on metaculus, polymarket, and kalshi
> Try another LLM call to sort the most important news.
>  
> Answer manipulation:
>  
> Test a multiple different LLMs at once and take median/mean
> Test multiple of the same model and take median/mean (Used last tournament)
> Statistical manipulation of predictions. Ex. reduce confidence of all predictions by 10% and remove outlier predictions.
>  
> Prompting:
>  
> I plan to improve upon the specialized prompts I used last competition:
> Test chain of thought prompting for specific question categories. Ex. for stock movement questions use Implied volatility and a random walk
> Test altering prompts with more general forecasting principles
>  
> Overall I will focus on selecting the best model, best combination of news sources, best prompting, and best answer manipulation in that order. I will also be sure each change I make is tested and provides meaningful improvement.

**Panshul42 (Fall 2025):**

> Architecturally, it will be similar to what I was doing in Q2, I'm planning to update models to GPT-5, and GPT-5-mini, as their forecasting performance seems a lot better than Claude models (for now). Also, will include computer use and additional information retrieval elements to more accurately fetch information from the resolution source.

**adjacentwire (Fall 2025):**

> At Adjacent we are building a News API and Data API which aggregates markets and comments from real-money, play-money, and forecasting platforms. So far Kalshi, Polymarket, and Metaculus, soon Limitless, Manifold, and Clarity markets.
>  
> In the last tournament for Q2 we forked the Metaculus template bot and only changed one thing, added in Adjacent API for markets in the "background" information given to the bot.
>  
> This was helpful enough to have us place #11 when not even starting at the beginning of the tournament.
>  
> We have improved our API and are looking to incorporate news, markets, and comments this round along with using a custom model built on top of Rei.
>  
> In the future we are considering using custom datasets and APIs for popular categories of questions.

**WobblyBotTeam (Fall 2025):**

> Initially, the bot will be more grounded on prediction markets forecasts and a few tried and true heuristics, that we've had success with in human tournaments, for the LLMs.
>  
> After the initial version, we plan on experimenting on various different heuristics and strategies to make the bot do more independent predictions.

**NashBot (Fall 2025)**

> I don't expect this bot to be successful but thought it would be worth experimenting to see if there is a new local maxima. It seems like quite a few bots average out their results and so I'm trying one that anchors its forecasts to base rates and takes one of the more extreme options. It first asks an optimist prompt to build a strong argument for an event happening. A second, sceptical prompt tries to find flaws in those arguments to produce a forecast closer to the base rate.

**Spagbot (Fall 2025):**

> I plan to use five main components: 1) prompt multiple models for an ensemble forecast using the same prompt but potentially different temperatures, 2) Bayesian approaches mixed with 3) for sociopolitical questions, game theoretical approaches, and 4) for all, monte carlo, 5) use bot performance to develop on-the-fly calibration instructions to be build back into the prompts.

**5cast-v1**

> Allow the model to operate freely outside the "assistant researcher / main researcher" framework by using an alternative that will focus on a few key changes: - Tell the model directly which context it is executed in (AI forecasting benchmark). - Also allow the model to decide on a forecasting approach for itself. - Provide links to analyses from past years (FutureEval Results Year 1), so the LLM may learn what worked and what didn't (? Experimental) - Unify research and forecast into one LLM, with research/asknews/exa as search I am banking on Opus and its agent capabilities, and the fact that other members won't pivot to it due to its relatively recent release date. In the Q2 Metaculus benchmark results, o3 ranked highly. If agentic capabilities (and using them correctly) and forecasting ability are positively correlated, the bot should do relatively well.