Experimental picks · Official ATS picks
How these forecasts work
Version betting-residual-v1 · experimental · no validated market edge.
publication
Owner-authorized immediate experimental publication; no validation or profit claim. Existing official ATS editions and scheduled publishing continue unchanged. New market forecasts have a separate record.
spread
Earlier-season fit: remove raw team-only scale/home bias versus market, then ridge-fit actual margin minus market-implied margin on centered disagreement. Ridge penalty 100 on standardized feature, no residual intercept. Compare market, raw 5%, raw 10%, raw 30%, conditional team/QB 30% and residual candidate.
moneyline
Ridge logistic adjustment to same-time no-vig moneyline log odds using standardized scale-neutral team-only disagreement; penalty 100, no intercept. Exclude ties from binary training and scores; report them. Live two-way probabilities conditional on no tie, ties void under recorded two-way settlement convention.
totals
Ridge regression of actual total minus market total on earlier-week scoring form minus market total. Form averages points scored plus allowed over each team's last eight games; minimum four games per team, previous season allowed. Penalty 100, no intercept. Historical scores updated only after the whole target week.
evaluation
Whole-season expanding walk-forward with at least 400 prior eligible rows. Fixed features/penalties; no search or promotion. Paired errors clustered by season/week with HAC lag 4 intervals. History already explored; revised data and unverified close vintages mean retrospective diagnostics only.
forward
Archive every game and exclusion at first observation each Eastern date and model revision. Freeze decision-time paired book prices, market baseline, inputs and model. Primary cohort Saturday 08-09 ET; Wednesday separate; remaining observations monitoring. Only first public forecast per event and market counts in public beta record. Changes never overwrite it.
clv
Opening/closing not claimed without verification. Track last observed pre-kickoff paired quote and gap to kickoff; comparisons labeled observed movement. No automatic promotion based solely on CLV or sample count.
futureCandidates
pregame confirmed QB scenarios, forecast wind and roof, rest and travel, injury timing, early-season priors
notModeledV1
explicit QB/player injury adjustments, weather point adjustments, travel, pace
Moneyline reliability, held-out seasons
| Probability range | Games | Mean forecast | Observed wins |
|---|---|---|---|
| 0–10% | 2 | 9.2% | 0.0% |
| 10–20% | 42 | 16.0% | 23.8% |
| 20–30% | 142 | 25.5% | 21.8% |
| 30–40% | 229 | 35.5% | 34.1% |
| 40–50% | 249 | 44.4% | 40.6% |
| 50–60% | 296 | 55.6% | 55.1% |
| 60–70% | 309 | 64.4% | 61.8% |
| 70–80% | 285 | 74.9% | 75.1% |
| 80–90% | 110 | 84.8% | 90.0% |
| 90–100% | 10 | 91.3% | 90.0% |
Small extreme-probability groups are especially uncertain. Ties are excluded from this binary diagnostic and void in the public two-way record.
Paired uncertainty
spread: weekly mean candidate loss minus market 0.04576; 95% interval [-0.18550, 0.27703]. Negative favors the candidate.
moneyline: weekly mean candidate loss minus market 0.00023; 95% interval [-0.00053, 0.00099]. Negative favors the candidate.
totals: weekly mean candidate loss minus market 0.04579; 95% interval [-0.87260, 0.96419]. Negative favors the candidate.
Intervals use weekly groups and HAC lag 4. They do not erase prior model-selection exposure.
Historical testing and limitations
Whole-season walk-forward. No candidate has established an advantage over the market. Historical quote timestamps are unverified; these are retrospective diagnostics, not executable Saturday returns.
ATS
| Forecast | Games | RMSE ↓ | MAE ↓ |
|---|---|---|---|
| market | 1679 | 12.641 | 9.759 |
| raw05 | 1679 | 12.641 | 9.754 |
| raw10 | 1679 | 12.644 | 9.751 |
| raw30 | 1679 | 12.681 | 9.772 |
| qb30 | 1679 | 12.671 | 9.767 |
| residual | 1679 | 12.643 | 9.758 |
Totals
| Forecast | Games | RMSE ↓ | MAE ↓ |
|---|---|---|---|
| market | 1737 | 13.048 | 10.305 |
| residual | 1737 | 13.051 | 10.305 |
Moneyline
1674 non-tied games. Brier score: market 0.2109, candidate 0.2110. Log loss: market 0.6092, candidate 0.6095. Lower is better.
All three paired uncertainty intervals include zero. Probability estimates remain experimental. Forecasts use current market prices as their baseline, a small fitted residual adjustment, and no explicit injury, weather or pace point adjustment in v1.
Entertainment only
Fantasy Overdrive provides sports analysis and picks for informational and entertainment purposes only. We do not accept bets, take wagers, or operate a sportsbook. Nothing on this site is financial or investment advice. Predictions are uncertain; we make no guarantees of accuracy, winnings, profit, or any particular result. Past results do not guarantee future performance.
If gambling is causing problems for you or someone you know, call or text the U.S. National Problem Gambling Helpline at 1-800-MY-RESET (1-800-697-3738), text the helpline, or chat online. Free, confidential support is available 24/7.