Methodology

Every open Kalshi and Polymarket contract settling within the next 120 days is pulled from each exchange's public market data: best bid, best ask, last price, spread, volume, open interest, order-book depth where published, and the settlement date and resolution source. Nothing is simulated or estimated. If a feed fails, the app says the data is not connected rather than showing stale numbers as live.

Buying YES costs the ask. Buying NO costs 100¢ minus the bid. The model always prices against what you could actually pay right now, never against the mid or the last trade, because that is where most naive edge estimates disappear.

The model starts at the mid price and applies three adjustments:

• Cross-exchange blend: when the same event is listed on both Kalshi and Polymarket, the other venue's mid price is blended in as a second independent opinion, weighted by relative traded volume and capped at 50% so the venue you would trade on always leads. Matching is automatic, based on title similarity and settlement date, and contracts can settle on different rules.
• Drift: the change between the oldest and newest third of the captured price history, capped at ±15¢ and weighted at 40%. Persistent directional flow carries information.
• Longshot adjustment: a bounded pull of up to ±4¢ toward 50¢, reflecting the well-documented tendency of event markets to misprice the tails.

The result is clamped to 1–99¢ and becomes the fair YES probability. Edge is that fair value minus the executable price on each side, minus the exchange's own trading fee on that side. The side with the larger fee-adjusted edge is the published signal, and the maximum entry price is set below fair value by one cent plus the fee, so you are never told to pay a price where costs swallow the edge.

Fees are modelled from each exchange's published schedule: Kalshi charges roughly 0.07 × price × (1 − price) per contract, which peaks near 1.75¢ around the 50¢ mid and falls away at the tails; Polymarket charges no per-trade fee on binary contracts, so a small flat allowance is applied for settlement friction instead. Every edge figure shown in the app — the calls, the scanner, the daily email and the public track record — is net of these costs.

Confidence is scored, not asserted. It rises with the size of the edge, tightens with the spread, and increases with volume and open interest. It is reduced when settlement is more than 120 days away, because capital sits idle and estimates drift. Low-confidence markets are excluded from the daily call of the day and from the public track record.

On demand, a language model reads the market's resolution rules, the price history and recent web reporting on the underlying event, then states what has actually happened and whether the question is already decided in the real world. This exists to catch markets the price has not caught up with. It can still be wrong, so it is presented with its sources and never overrides your judgement.

Allocation is a quarter-Kelly fraction of a hypothetical model bankroll, capped at 25%. Full Kelly is (q − p) ÷ (1 − p), where q is the model's probability and p the entry price. Quarter Kelly sacrifices growth for a large reduction in drawdown.

This figure is deliberately expressed as a percentage, identical for every subscriber. It is not based on your finances, and Pricedger does not ask for them.

• Edge is calculated before exchange fees, slippage and taxes.
• The quant model has no knowledge of news on its own; only the AI deep dive reads the world.
• Thin markets can show large phantom edges; liquidity filters reduce but do not eliminate this.
• Cross-exchange pairs are matched by title similarity and can be wrong — always read both sets of rules.
• A tight spread with tiny size is not the same as real liquidity.

Up to ten of the highest-conviction calls are written down publicly each day, before the event, with the price the model would have paid. When the market settles, the exchange's own outcome decides whether the call was right. Hit rate, Brier score, log loss and a calibration curve are published on the track record page. Nothing is backfilled and losing calls are never removed.