Model methodology

Schedules and recent results come from sports feeds; bookmaker quotes carry the provider, bookmaker and capture time. Missing or stale data prevents a pick. Social references are unverified context and do not change probabilities.

The current engine starts from the consensus of several bookmakers: the margin is removed to get a fair price, which is then blended with a Poisson model of recent team form. A pick needs positive expected value at the best available price. Before publication an AI research step looks for injuries, lineups and other news with cited sources; it can only keep, downgrade or veto a pick, never change its price or probability. Confirmed lineups and individual starting-pitcher models are not part of the model. Profitability is not guaranteed. NHL, MLB and MLS are supported; the broader match calendar is separate.

For a binary bet: implied probability = 1 / decimal odds; edge = 100 × (model probability − implied probability); EV = probability × odds − 1. At 68% and odds 1.35, implied probability is 74.07%, edge −6.07 percentage points and EV −8.2%. This is not positive value. Push and split markets require different payout calculations.

Predictions and their source snapshots are saved before kickoff, then matched with final outcomes. Calibration uses earlier training data and later test data, with Brier score and log loss compared against simple baselines. A minimum sample is only a starting point. An improvement is reviewed before any model promotion.

For each league and market, the engine can adjust the blend of model estimates and fair bookmaker probabilities after at least 50 valid settled predictions. Repeated snapshots count once; post-start predictions and unknown leagues are excluded. Learned model weights are bounded between 5% and 40%. A negative closing-line drift signal needs at least 30 valid pre-match price captures before reducing the model weight. Strong historical teams remain observations, not proof of future success or automatic model promotion.

Scheduled research: 11:20 / 15:20 / 19:20; publication: 12:00 / 16:00 / 20:00 (Europe/Prague). At most 15 analyses per day, 5 per league; fewer or none may qualify. Daytime editions cover matches until 20:45; the evening edition covers 20:45 until 12:45 the next day. Each run requires at least 45 minutes before kickoff. Results and model review run at 11:00. Preseason is excluded from headline results. Quotes are checked when an analysis is created; later price and lineup changes are not monitored continuously.

All published wins, losses and voids remain visible. Fixed units are a model portfolio, not proof of money wagered. The product has not demonstrated independent profitability. Previous losses are not erased by a weekly reset.

Educational articles are AI-assisted and clearly labelled. A different AI model reviews the draft, recalculates the examples and checks claims against the supplied data in both languages. Structural checks and monthly spending limits run before publication. This automated review does not replace a human fact-check. Sources, limitations and correction contact remain visible. No fabricated interviews or claims of guaranteed returns.