Methodology

Pricedger continuously ingests public market data from Kalshi, Polymarket US and Polymarket Global. The standard active research catalog covers binary contracts settling within the next 365 days. Combination and parlay-style Kalshi contracts are excluded from automated scoring. Closed, expired and resolved listings are archived to preserve historical integrity rather than presented as active.

Universal on-demand ingestion: in addition to the pre-indexed catalog, Pricedger supports on-demand resolution for publicly listed Kalshi events and Polymarket markets, including sports spreads and game lines. Pasting an unindexed exchange URL queries the venue's public endpoint, unarchives or stores the contract and generates a standard research page. Rate limits apply to keep this fair for everyone.

Custom and off-platform contracts: for propositions outside Kalshi and Polymarket — for example PredictIt, Manifold or a private agreement — users can initialize a custom research contract by specifying the question, the current market price and the resolution conditions. Custom contracts run through the quantitative baseline, quarter-Kelly sizing and AI scenario protocols identically to exchange-listed markets. They do not receive automated price polling and do not enter the public track record.

The available fields differ by venue. They can include the best bid and ask, last or indicative price, spread, traded volume, open interest, displayed liquidity, settlement date, resolution rules and recent price history. Official hourly bars and top-of-book size are used where the venue and workflow provide them. Polymarket US does not publish the same activity fields as the other feeds; Pricedger leaves those values unavailable rather than inventing them.

Data is refreshed on a schedule, but no feed is guaranteed to be continuous or current to the second. If required data is missing, Pricedger lowers confidence, omits the affected input or declines to price the contract. It does not label an unavailable feed as live.

The model does not treat the last trade as an available entry. Buying YES is evaluated at the current YES ask. Buying NO is evaluated at 100¢ minus the current YES bid. A market without a two-sided quote receives no quantitative side call.

Edge is calculated after the estimated entry fee. Kalshi is modelled with its published coefficient of 0.07 × price × (1 − price) per contract. Polymarket US uses 0.0695 in the same price-based form. Global Polymarket is modelled with no central-order-book trading fee. These are per-contract estimates; venue rounding, market-specific terms, rebates or later fee changes can make the amount charged differ.

Where the venue publishes a resting order book, Pricedger also sweeps the displayed levels to estimate average fills for standardized $100 and $500 orders. It labels execution risk as low, medium or high. A high-risk book lowers confidence, and the maximum entry is reduced by the estimated $100-order slippage, within a bounded allowance. The paper-sizing display uses that realistic $100 fill when available. The underlying edge comparison still begins at the best executable quote, so the depth readout and lower price limit should be considered together.

Full depth is not available in every workflow or from every venue. In particular, Pricedger does not currently read the Global Polymarket order book. Hidden orders, cancellations and fast price changes can also make an actual fill worse than the displayed sweep.

The starting point is the midpoint of the current YES bid and ask. Pricedger then applies the following bounded adjustments before comparing fair value with the executable price:

• Cross-venue price: when a likely counterpart is found on another venue, its midpoint is blended in by relative reported volume, with the venue you can actually trade always keeping the leading weight, then reduced by both the text-match and contract-equivalence scores. When activity data is absent, the two books receive equal starting weight before those adjustments.
• Captured drift: the difference between the oldest and newest thirds of stored YES-price observations is bounded, so a short or thin history can never swing the estimate on its own. It carries the larger share of the trend read when official candles are unavailable.
• Hourly momentum: with at least six official bars, Pricedger compares recency- and volume-weighted early and late prices over a recent trading window. Candle-based momentum leads the coarser snapshot drift when both are available, and the two are combined rather than one overriding the other.
• Order-book pressure: when best-bid and best-ask sizes are available, a micro-price estimates which side has more resting pressure. A bounded share of that skew enters fair value, so a single lopsided quote cannot drag the estimate.
• Longshot adjustment (favourite–longshot bias): fair value is adjusted for volume-weighted order-book momentum and short-term drift, bounded by category-specific calibration curves that account for retail lottery premiums in political and longshot contracts. Cheap contracts carry that premium, so the blended midpoint of a contract below 50¢ receives a haircut and one above 50¢ a small lift — never a pull of a longshot toward even. The curves are steepest in retail-driven politics and culture books, moderate in geopolitics and crypto, lighter in sports, and close to nil in economic data and weather, where prices already reflect the bias. Each curve is capped in cents, so no category adjustment can move an estimate dramatically. The category is taken from the listing, or inferred from the ticker series and title. These are conservative priors, not universal empirical laws. Only calls with a model win probability of at least 30% are logged to the public track record.

The raw estimate is bounded to 1–99%. The model calculates net edge for both sides: fair payoff probability minus the executable price and estimated fee. It selects the side with the larger net edge, but separately warns when neither side clears the platform's minimum worthwhile-edge threshold.

Maximum entry sits below the called side's fair value by a fixed minimum margin plus the estimated fee. Choppy markets earn extra protection: when hourly volatility is available, a buffer that scales with the typical hour-to-hour move is deducted, and where order-book depth is available a further bounded allowance is deducted for the estimated slippage on a standardized $100 order. These margins are conservative and frozen within a model version, so the same inputs always produce the same ceiling.

Prediction markets are prone to behavioural distortions, narrative momentum and liquidity gaps. Rather than relying on a single forecasting technique, Pricedger synthesizes several related (not fully independent) probability estimates through a Visible Ensemble:

• Quantitative baseline: order-book midpoints adjusted for drift, volume-weighted momentum, micro-price order-book pressure and category longshot bias (Section 3).
• Empirical base rate: outside-view reference-class historical frequencies.
• Parametric model: objective probability distributions derived from official government or weather forecast data where applicable (Section 9).
• Structured AI superforecasting estimate: synthesis of resolution criteria, recent domain news and fundamental catalyst scenarios (Section 10).

Consensus calculation: when two estimates are present, the consensus is their arithmetic mean. When three or more estimates are available, the ensemble takes the median of the valid estimates. This prevents an extreme outlier from pulling the aggregate estimate away from the consensus. The rule is deterministic, so the same inputs always produce the same displayed number.

Ensemble disagreement spread: the model tracks the absolute spread between the highest and lowest valid estimates. Small spreads are treated as aligned, moderate spreads as mixed, and wide spreads are flagged as sharp disagreement. The flag is a visible warning, not an automatic change to the paper allocation: users are alerted to look for underlying regime changes or ambiguous resolution criteria before relying on the call.

Kalshi and Polymarket prices are not independent observations: traders, news and arbitrage can transmit the same information between venues. Pricedger therefore treats another book as corroborating market evidence, not as a separate vote.

The initial matcher compares normalized question and outcome wording, requires a minimum similarity score and normally limits matches to settlement dates close together. It prefers the higher-volume counterpart when matches are otherwise close. Each candidate pair then receives a 0–100 contract-equivalence score. That score compares named numeric thresholds, settlement timing, the official resolution source, stated time zone, delay or cancellation treatment, and wording similarity.

Higher scores are labelled “Same contract,” mid-range scores “Likely same,” and lower scores “Basis risk.” The score proportionally reduces the other venue's weight, and a pairing in the lowest band is not blended at all. A sharp price disagreement lowers confidence when the pair otherwise remains usable. Cross-venue locked-price displays also require executable quotes on both legs; a last or indicative price is not treated as an available arbitrage entry.

This remains a rules-based comparison, not a legal determination of equivalence. Source wording can be incomplete or difficult to parse, so users should still read both contracts.

Whale Watch scanner: Pricedger tracks large-notional transactions across Kalshi and Polymarket Global to monitor large-notional positioning. Trades are filtered into tiers starting at $5,000, scaling through $10,000, $25,000 and $50,000+ notional. Free research accounts view tier-1 trades ($5,000+) delayed by 60 minutes; Pricedger Pro subscribers access all tiers as they are collected, roughly every five minutes.

Top Wallets leaderboard: on Polymarket Global, public blockchain wallet addresses are aggregated to display historical trading volume, realized profit and loss, win rate and active holdings.

Smart Money Divergence: for qualifying markets with large-trade flow, Pricedger evaluates net 7-day whale volume (YES versus NO buying) against the contract's 7-day price move. When whale flow heavily buys one side while the spot price is flat or moving in the opposite direction, an accumulation/divergence signal is flagged. When whale block buying and market price move in tandem, the trend is marked as whale-backed.

Whale Watch records are derived from public venue feeds and blockchain settlement logs. They do not account for off-book hedging, multi-leg delta-neutral positions, or split orders executed below the tracking threshold.

Confidence is a rules-based label, not a probability that Pricedger is correct. The score rises with larger fee-adjusted edge, tighter spread, and stronger reported volume and open interest. Cross-venue agreement and momentum that supports the selected side can raise it. Sharp cross-venue disagreement, momentum against the call, high hourly volatility and a settlement date far beyond the usual research horizon can lower it. A partially fillable or high-slippage $100 order also lowers the score where depth is available.

These labels are used for research triage. They are not yet fitted confidence intervals, and open interest is treated as an execution-quality signal rather than proof of forecast accuracy. Low-confidence markets are excluded from the call of the day and the automatic public record.

Every resolved public call is scored on the side Pricedger selected. The public track record reports hit rate, Brier score, log loss, confidence-band results and a reliability curve comparing average stated probability with actual frequency. Category tables include sample sizes so thin evidence is visible.

The same record feeds a conservative correction into new quantitative estimates. Once a category has a meaningful sample of settled calls, Pricedger measures average stated confidence minus realized hit rate. When that gap shows overconfidence, the new estimate is pulled back toward even by a multiple of the historical gap, capped so no category correction can flatten a call. Categories with too little history fall back to the overall record. Underconfidence never pushes a new forecast farther from even.

This is a transparent feedback rule, not a full out-of-sample calibration model. The sample is still shaped by which calls qualified for publication, and results can change as more markets settle. See the public track record for the current evidence.

For supported fast-settling markets, the Pro AI deep dive adds a structured feed from the agency most closely tied to the outcome. CPI, core CPI, unemployment and nonfarm-payroll inputs come from the U.S. Bureau of Labor Statistics. The current effective federal-funds rate and target range come from the Federal Reserve Bank of New York. Supported city-high temperature markets use National Weather Service forecasts for the relevant settlement location.

The economic feeds summarize the latest reading, recent changes and observed variability. The weather feed supplies up to four daytime forecast periods for supported stations. The deep dive is instructed to compare the contract threshold with those readings, trends and typical swings before applying news-based updates. It displays the source and “as of” date with the evidence.

For supported CPI, core CPI, unemployment, payroll and city-temperature questions, a parametric layer can turn that baseline into a direct probability of YES. It first parses a recognizable threshold such as “above 0.3%,” “below 4.5%” or “between 72° and 73°.” A normal distribution then compares the threshold with the expected reading and its recent or forecast uncertainty. Temperature ranges receive a small continuity adjustment so a narrow band is neither over- nor under-weighted. The result is bounded well clear of 0% and 100% so a simplified distribution never claims certainty.

Inflation thresholds are matched to the measurement basis the wording implies — larger thresholds read as year-over-year, smaller ones as month-over-month. Weather uses the available daytime forecast distribution. These are explicit modelling assumptions, not facts about every contract. If the wording is not recognized, the threshold is implausibly far from the feed, or the required series is unavailable, the model returns no statistical probability rather than guessing. Smooth normal modelling is deliberately not used for stepwise Fed decisions.

Presidential-approval and national congressional-control research can also use independent polling averages from USPollingData.com. Approval thresholds can receive a parametric baseline around the current average. House and Senate research receives the generic-ballot margin, approval average and reported Senate toss-up count as context, but no automatic control probability because national averages do not uniquely determine individual races. This source is independent polling aggregation, not an official government feed.

Government results are cached for three hours and polling results for six. Retrieval is best-effort: when a source is unavailable, has too little history, changes format or does not cover the market, the deep dive continues without a structured feed rather than fabricating one. USPollingData currently provides national figures, not state or district race models, and its definition may differ from the pollster or average named in a contract. These feeds and probabilities currently inform Pro AI deep dives; they do not directly change every quantitative call, scanner score or call of the day.

A Pro deep dive combines the contract rules, exchange data, quantitative baseline, available hourly bars, a likely cross-venue counterpart, any applicable official domain feed and cached recent web reporting. The language model must work through a structured protocol rather than jump directly to a recommendation:

• Start with an outside-view reference class and a historical YES base rate. Where an official feed applies, compare the threshold with the latest reading, recent trend and typical swing.
• Update incrementally for market-specific evidence and identify when the real-world event may already be decided.
• Present the strongest case for YES and the strongest case for NO, with the final estimate between those scenarios.
• Check the official resolution source, deadline, timezone, postponement, cancellation and ambiguity risks.
• State concrete observable conditions that would invalidate the call, and return no call when the question is not responsibly priceable or lacks meaningful edge.

The deep dive reports three explicit probability scenarios in the Scenario Matrix: a Base Case (the most probable trajectory given current fundamentals and base rates), a Bull Case (the thesis and probability of YES assuming positive catalysts or upside revisions materialize) and a Bear Case (the thesis and probability of NO assuming negative catalysts or counter-evidence materialize). Every scenario probability is stored as probability of YES; the NO card displays its complement so the two cases are directly comparable.

Each deep dive also outputs concrete, testable invalidation triggers — specific real-world developments that would break the thesis. Forecast revisions are logged in a timestamped trajectory, so users can see how the estimate adapted as new information arrived rather than only the latest number.

Deep-dive queries are cached across all users for three hours to keep analysis consistent and prevent repetitive model drift. Web reports and model output can be incomplete, stale or wrong. The AI layer does not override contract rules, does not place orders and is not silently substituted for missing exchange data. Its estimate is a research opinion rather than a verified outcome.

Position sizing uses quarter Kelly for a binary contract. If q is the estimated probability of the called side and p is its entry price as a decimal, full Kelly is (q − p) ÷ (1 − p). Pricedger uses one quarter of that fraction, floors negative values at zero and caps the displayed allocation at 25% of a hypothetical model bankroll.

The suggestion is then trimmed against open paper positions. One inferred market type may use at most 25% of the paper bankroll, and contracts grouped under the same event may use at most 10%. If either limit has no room left, no additional paper stake is suggested. These are concentration guardrails, not an estimate of statistical correlation.

This is standardized paper sizing, not individualized bankroll advice. It does not account for personal finances, correlations between different event codes, taxes or the possibility that the probability and liquidity estimates are wrong.

Every contract with published rules is audited by a deterministic forensic engine. The same rule text always produces the same result, and the audit uses no AI credits. It classifies the settlement source as an official agency or league, a news-consensus standard, a decentralized token-holder vote (UMA) or unspecified, and it detects inverted polarity, where the event happening resolves NO.

• Source risk: subjective “consensus of credible reporting” language and decentralized settlement are flagged because they are open to interpretation and dispute.
• Timing: confirmation or dispute windows, non-US-Eastern cutoffs and recount or certification clauses that can delay settlement.
• Scope: explicit carve-outs that do not count, narrow geographic definitions, and whether an announcement alone suffices or a formal completed action is required.
• Postponement and cancellation: rules that change when a contract settles, including 50/50 settlement on cancellation.

Each flag carries a fixed weight. Weights are summed, adjusted slightly for rule length and for a clear official source, and bounded to a 0–100 dispute-risk score: under 20 is low, 20–44 medium and 45 or more high. The score is a heuristic ranking of wording risk, not a calibrated probability that a dispute will occur. When an AI Deep Dive runs, the same single request also returns a plain-English explanation and trading takeaway for each flagged clause; no extra AI call is made.

Pro users can configure up to three paper agents with a starting paper bankroll, a Kelly fraction, minimum net edge and confidence, maximum days to settlement, venues, per-trade size and exposure caps. Agents are numbers-only: they use the same deterministic side call and fee-aware edge described above and make no AI calls.

An hourly server runner screens the active catalog for each agent, keeps only markets that pass every configured filter, and records simulated fills at the executable entry price, sized by fractional Kelly and trimmed by the agent's caps and remaining paper cash. When the exchange settles a market, the runner marks the position won or lost and books P&L. Users can change agent settings but cannot write or edit agent trades, so each agent's ledger is tamper-proof. Strict filters can mean an agent places no trades for long periods. Simulated fills ignore queue position and partial fills, so agent results are not live-trading results.

Global Polymarket trades are attributed to public wallet addresses. Pricedger groups wallets that repeatedly take the same side of the same contracts within a tight, bounded time window and across more than one contract. The strongest recurring patterns are shown as possible syndicates with their shared contracts and combined size.

Pro users can follow individual wallets and receive in-app alerts, plus optional hourly email digests, when a followed wallet places a new trade of $5,000 or more. Clustering indicates coordinated timing, not shared ownership or intent; unrelated traders reacting to the same news can cluster. Kalshi and Polymarket US do not publish trader identities, so clustering covers Global Polymarket only.

On Analyze Any Contract, users can drop, browse or paste a screenshot of a prediction-market contract. The image is resized in the browser and sent in a single vision request that extracts only visible fields: venue, question, specific outcome, any visible URL, YES price, settlement date and rule text. Text inside the image is treated as data, never as instructions.

Matching is then deterministic. A visible Kalshi or Polymarket address is resolved exactly like a pasted link. Otherwise the question and outcome are compared separately against contract titles and subtitles, ranked by keyword overlap, detected venue and closeness of the visible price to the market price. A single clear match opens the full live terminal; several close matches are listed to choose from; with no match, the extracted fields prefill a custom contract. Daily screenshot limits apply: 30 for paid Pro and 5 for trial and free accounts. Extraction can misread low-quality or cropped images, so the matched contract should always be confirmed.

Up to ten eligible high-conviction calls are recorded automatically each day before the event, including the venue, side, executable entry, estimated probability, edge and model confidence. The exchange's own outcome later determines whether the call was correct. Records are not backfilled, and losing calls are not removed.

The published record is a selected record of qualifying calls, not every market scored by the system. This selection can make it unsuitable for claims about the accuracy of the entire market universe. Brier score and log loss are more informative than hit rate alone because they penalize confident mistakes.

Each call carries the model version that produced it. Calls made before version tracking are labelled “legacy”; new calls use v3.0. Each row on the Track Record shows its model version badge so materially different forecasting systems are not silently pooled. Prototype sandbox calls logged on 20 September 2026, before the public record began, were removed and are not part of the published record.

• Custom contracts: user-created propositions do not receive automated price polling or live settlement verification; their pricing and sizing reflect the inputs provided by the user.
• Whale flow completeness: Kalshi trade history is sampled from public transaction streams and can omit fills during rapid order-book bursts. Polymarket US does not publish a public real-time trade tape. Whale Watch data should be treated as high-conviction sampling, not a complete audit of all exchange activity.
• Ensemble weighting: the median/mean consensus is deterministic and bounded. It does not dynamically adjust model weights by past accuracy on a per-subtopic level.
• Several coefficients — including drift, momentum, longshot pull, confidence points and calibration shrink — are bounded expert-designed heuristics. They have not yet been estimated and frozen through a formal walk-forward, out-of-sample process.
• The execution model sweeps visible resting levels for standardized $100 and $500 orders; it does not predict hidden liquidity, cancellations, queue position or the impact of a live order. Global Polymarket depth is not yet read.
• Contract equivalence is a heuristic score over recognizable thresholds, dates, named sources, time zones, cancellation language and wording. It does not understand every legal clause and can miss a material basis risk.
• Market data coverage and field quality vary by venue; missing values reduce what the model can infer.
• Category-level calibration uses a coarse grouping and a minimum sample size. It does not yet condition jointly on venue, horizon, probability band or regime.
• The AI research layer can hallucinate, misread a source or choose a weak reference class. Citations and settlement rules should be checked directly.
• Structured domain feeds currently cover selected U.S. inflation, labour, federal-funds, city-temperature, presidential-approval and national congressional-control markets only. Government figures and independent polling are cached, can be revised or unavailable, and currently inform the AI deep dive rather than every quantitative call.
• Parametric probabilities use simplified normal distributions and strict wording rules. They can misidentify the intended measurement basis or date, and intentionally return no estimate when the contract cannot be parsed safely.
• Suggested paper stakes are trimmed so open positions in one market type stay under 25% of the paper bankroll and one event under 10%. These caps group by inferred market type and event code only; they do not model correlations across different events.
• Paper results do not include realistic fill uncertainty and cannot be assumed to equal live-trading results.

The next statistical priorities are broader race-level and international structured feeds, deeper venue coverage and walk-forward validation within frozen model versions. Until those exist, Pricedger's output should be read as a transparent research estimate, not a promise that the market is wrong.

Pricedger is for research and paper trading only. It cannot place real orders, withdraw funds or move money. Nothing on the platform is financial, legal or tax advice. Prediction market availability varies by jurisdiction, and users are responsible for checking the rules that apply to them.

Not affiliated with Kalshi or Polymarket.