The Method

The baseline is the sportsbook's implied probability after removing its margin from both sides of the same bout. Model agreement and confident analysis are candidate signals; they are not proof that the market is wrong. A sportsbook quote is also not an executable Kalshi price.

Compare every candidate with that baseline on the same eligible bouts using Brier score and log loss, where lower is better. A coin flip scores 0.250 on Brier, but beating a coin flip is insufficient when the market already forecasts the favorites well.

Join predictions, odds and settled outcomes by event and fighter pair. Keep a single eligible prediction snapshot, validate both sides of the odds, and report excluded or ambiguous rows. Exclude snapshots that cannot be established as pre-event and bouts without a usable result. A missing value stays missing.

UFCStats, opponent history, ratings, age and layoff can help only if their values were available before the forecast. Reconstructing an old fight with today's career totals leaks its outcome into the features. Date-only timestamps cannot establish the exact information available at a particular betting time; the report must retain that limitation.

The research pipeline compares the unchanged market forecast with shrinkage blends and a regularized logistic residual model. The residual model starts from the market's log odds and learns a limited correction from the available prediction signals. Regularization penalizes large corrections so a small sample has less opportunity to manufacture an edge.

Blending weights and penalty strength are selected using earlier cards inside each training window. The market remains a valid winner of model selection. Language-model confidence is never accepted as a measured win probability just because the explanation sounds persuasive.

Validation moves forward through time. Train on earlier event dates, select settings inside that earlier history, and score the next unseen event date. Keep every bout and prediction from an event together. Leaving one event out while training on later events is not a simulation of forecasting the future.

Report the sample size, validation folds, model and market scores, and event-level uncertainty in their difference. A favorable point estimate with a confidence interval spanning zero does not establish improvement. Ranking many settings on the same history produces a hindsight winner, not an independently verified strategy.

For a contract paying $1, the break-even probability is its purchase price plus fees per contract. A trade also needs a margin for uncertainty and execution costs. Fresh quotes, matching outcomes, sufficient liquidity, event timing and available bankroll all matter. A positive model-minus-price number alone is insufficient.

The trading desk shows the estimated debit including fees, spending limits, readiness checks and reasons to abstain. Paper plans help collect evidence; they do not prove that orders would fill. Live automation requires verified research and execution readiness. Manual trading has its own order review and current exchange fee estimate. Actual fees and rounding are determined by Kalshi.

Updated 2026-09-19T22:40:02.145Z · research_only

No proven trading edge. Use paper tracking while collecting timestamped prices, fills and actual fees.

77 eligible bouts across 9 event dates. 2024-12-07 to 2026-09-19. 1173 source rows excluded by the audit. 1 chronological validation folds.

Live automation readiness

• Operator override: KALSHI_LIVE_AUTOBET is set, so live orders are sent for research, claude, gpt, consensus, double_lock, book without validated execution evidence.
• Historical sportsbook probabilities are not executable Kalshi prices.
• No independently validated, fee-adjusted forward Kalshi execution study has been promoted.
• Fewer than 200 held-out bouts across 20 events.
• Held-out probability improvement over the market is not statistically established.

8595 eligible bouts, split chronologically into 5286 training, 1309 validation and 2000 untouched test fights. Model selected on validation: boosting.

Outcome prediction experiment, not a Kalshi profit backtest. Model selection was frozen before the final holdout.

The market comparison: 66 shared bouts across 7 event dates.

Same held-out bouts with independently verified pre-event sportsbook prices. Small sample; no Kalshi fill or profit evidence.

Prediction accuracy: graded forecasts, calibration and performance relative to the market. Sportsbook proxy simulations: hypothetical trades using historical sportsbook probabilities and an assumed fee model. Kalshi execution: actual quotes, orders, partial fills, fees and settled returns in the account.

The Record tab labels historical simulations and research separately. Your History tab shows paper and live activity separately. No historical accuracy score or proxy profit establishes routine future profitability.

Open the trading desk