How eagL quantifies prediction markets

Pricing model for crypto strikes

eagL prices Kalshi crypto strikes and brackets under geometric Brownian motion. Drift is estimated per minute from stored minute candles. Variance uses an exponentially weighted moving average with a twenty-period half-life, so recent volatility counts more than last week's. Strike probabilities are read off the implied lognormal distribution. Past a two-hour horizon the model drops drift entirely, because at that distance the drift estimate is noise and keeping it would only add error with a confident face.

Forecast scoring

Once a market resolves, eagL scores the forecast it made: Brier score, log loss, and a ten-bin calibration curve that shows whether events the model called 70% actually happened about 70% of the time. The scores accumulate in the app, so you can watch the model earn its keep or fail to.

Order-book microstructure

Over the top eight levels a side, eagL computes microprice, depth imbalance, queue imbalance, liquidity elasticity and spread in basis points. These are the numbers a desk looks at to decide whether the displayed price means anything, and they update from the live book on your own connection.

Fees and sizing

Every venue charges differently, so eagL charges each venue's own formula rather than a flat haircut. Kalshi's fee schedule is applied and rounded up to the cent, the way Kalshi rounds. PredictIt takes its cut of profit at settlement. Coinbase is charged at 120 basis points taker. Fees apply at the fill VWAP, not the screen price: the size solver walks the ladder level by level and grows the position until the next contract stops paying.

Cross-venue matching

Matching the same event across venues uses IDF-weighted title matching behind a numeric gate, so a $100k strike never pairs with a $150k one just because the words agree. Play-money venues are dropped before matching starts.

Reachability replay

A cross-venue trade needs both legs to fill. eagL replays stored quotes thirty minutes forward from detection, aligns the two venues only where samples land within sixty seconds of each other, and counts the pair as takeable only if it still cleared all costs thirty seconds after detection. More on why this matters on the arbitrage page.

Validation

Backtests split chronologically, 60% for tuning and 40% held out, never randomly. Parameters are tuned on the early window, the held-out window is touched exactly once, and both numbers are shown along with the drop between them. The parameter sweep is capped at a few hundred combinations, because past that point the winner is luck. The backtesting page expands on this.

Minimum-sample gates

Before publishing any statistic, eagL requires 100 detections, 500 quote frames, 50 paper fills, 100 settled outcomes and seven days of observation. Below any floor it discards the figure instead of printing it with a footnote.

Known limitations

• Quote history is top of book, so the reachability replay proves the price held, not that the full size was there.
• The weather terminal reads settlement data. It does not price weather contracts, and we say so rather than shipping a weak model. See weather markets.
• Markets with no order book are priced from one assumed level and labeled as assumed.
• GBM is a workhorse, not the truth. Fat tails exist; the scoring exists to catch them.