MODEL 8.4 · ATS RESEARCH ADDED
See what the model expected, what happened, and what it learns for the next slate.
Paper trial · Model profitability unproven
Loading saved research. Fresh odds are not connected.
A lean chooses the side where the model adds a higher chance than the market reference. The record includes every directional lean, even when its price fails the value test.
Average miss in the winning margin. Lower is better. Both numbers use the same games with reference lines.
A good winner prediction can still lose against the spread. The score forecast and cover model are tested separately.
OPEN ANY NUMBER TO SEE THE REASON
BEATING THE SPREAD
ATS means against the spread. Winning the game and winning the bet are different.
We added recent ATS results, points above or below the spread, and this game situation. The change failed the existing test rule and is not used in the active model.
TESTING THE MODEL ON FUTURE GAMES
0 comparable closing prices saved. Real money recommendations: 0u.
Paper stakes come from the model and a bankroll. One unit is 1% of bankroll. The 5% per bet limit means the final size cannot exceed 5u. The trial starts at a 2% value threshold and allows for refunds.
These estimates compare a win rate with 52.38%, the break even rate at −110. The rough count is not a promise. The 80% power column allows for the chance that a real improvement is missed. Games in the same week can raise the sample needed.
Closing price evidence and win rates are monitored weekly. The success test happens once at the scheduled review.
2024, 2025 and 2026 are retrospective tests. These records include every directional lean, including prices that would not qualify. They are not the record of recommended bets. Consensus references are included; pushes are excluded from win rates.
ALL SUPPORTED SEASONS
These new 2022 to 2026 evaluations use the same search choices with fits beginning in 2017. They are retrospective research. The saved game forecasts below remain from 8.2. No old result counts as a new paper call.
Probability error, also called Brier error, measures how far the predicted chance missed the result. Log loss gives extra weight to confident mistakes. Lower is better for both. A negative model minus market difference favors the model. A range that crosses zero leaves the improvement uncertain. The slope asks whether a disagreement with the book helped predict the book’s error. The t value compares that slope with its uncertainty. These are tests of the forecast, not promises of profit.
The eligible archive has 570 games in 2020. The comparison removes that season from team updates, model fitting and similar games. For the 2022 comparison, 2019 replaces 2020 as the tuning year. Both results stay visible.
A stronger estimated advantage can produce a larger stake. We use a quarter of the Kelly calculation, reduce it for refunds, then apply the limits: 5% per bet, 5% per game and 20% for the week. Rounding never breaks a limit. A moneyline replaces the spread; it does not add another side bet.
This calculator does not change the fixed bankroll or records of the paper trial.
At 1u = 1% of bankroll, 5u uses 5% of the bankroll. The maximum paper recommendation is 5u. Real money stays at 0u.
Team ratings freeze before the week. Quote times stay attached to their prices.
Completed games update offense, defense and kicking. Missing play data stays marked missing.
After a complete week settles, the fixed procedure updates its parameters once. Every decision keeps its original parameters. The 8.2 comparison is logged without changing the selected model.
A spread meeting the fixed threshold receives a capped paper stake after all source checks. The first prospective trial covers spreads. Moneyline and team total models need separate validation. Real money stays at 0u.
Prices and decisions will appear only after they are received.
Each change must beat the market in every completed test season, with a clear combined result. We keep failed tests visible.
DATA CONNECTION
Saved results are available. Live collection needs setup.