Methodology · Honest Model Disclosure
SealedSlate does not claim to beat the market (closing odds). The backtest below is the proof. What we sell is not “winning picks” but a track record that is disclosed honestly, sealed with SHA-256 before each game, and reproducible and verifiable by anyone. Every number on this page is regenerated from the ledger by script.
Measured backtest EPL 1,350 games · walk-forward · pre-registered criteria
| Metric | Result | Interpretation |
|---|---|---|
| Value-bet ROI | -6.77% (n=1,198) | Expected-value betting loses — it does not beat the market |
| Model log loss | 0.9771 | Lower is more accurate |
| Market (Pinnacle close) log loss | 0.9497 | The market holds a significant edge |
Conclusion: the model is no more accurate than the closing odds (0.9771 vs 0.9497). That is why we do not sell “beating the market.” Instead, we disclose exactly and honestly “how close we come to the market.”
Calibration Predicted probability vs actual hit rate · Settled N=93
| Probability bucket | Predicted avg | Actual hit rate | Sample N |
|---|---|---|---|
| 30–40% | 40% | 100% | 1/1 |
| 40–50% | 44% | 33% | 1/3 |
| 50–60% | 55% | 56% | 24/43 |
| 60–70% | 65% | 68% | 15/22 |
| 70–80% | 73% | 76% | 16/21 |
| 80–90% | 83% | 67% | 2/3 |
“When we say 60%, what share actually hit?” — the sample is still small (N=93). As games accrue, each bucket's reliability improves. We never seal retroactively, so past gaps are left as gaps.
Accuracy metrics Based on settled predictions
| Metric | Value | Reference |
|---|---|---|
| Brier score | 0.222 | 0 = perfect · 0.25 = coin flip · lower is more accurate |
| Log-loss | 0.636 | Lower is more accurate |
| Hit rate (settled) | 63.4% | Favorite side · N=93 |
CLV · vs closing odds closing line value
CLV (model probability vs the market's closing implied probability) requires real closing odds to compute. Odds collection is not yet wired up, so CLV is on hold — we do not fill it in with approximations (honesty = the core of the product). Once an odds source is connected, we will disclose per-prediction CLV and the aggregate average with their signs intact.
Model features Only structured metrics that feed the probability
Home Advantage · season form (record) · home/away splits · offensive/defensive efficiency (ORtg/DRtg · score differential) · effective shooting % · turnover rate and other validated structured metrics have their log-odds contributions accumulated additively (every factor's %p is fully disclosed on each game page). Qualitative and folklore factors get a label only and are not reflected in the probability.
★Each factor's %p is a log-odds contribution converted into a probability, so it is non-linear — on a game page's waterfall, the plain sum of factor %p may not exactly equal (final − baseline) (a property of the logit transform). Displayed values are cross-checked and verified against the model's actual output.
★Each factor's %p is a log-odds contribution converted into a probability, so it is non-linear — on a game page's waterfall, the plain sum of factor %p may not exactly equal (final − baseline) (a property of the logit transform). Displayed values are cross-checked and verified against the model's actual output.
Limitations Honest disclosure
· No more accurate than the market (closing odds) — see the backtest above.
· The live track record sample is still small (settled N=93) — calibration sharpens as the sample grows.
· Closing odds are not yet collected → CLV and edge are on hold.
· Qualitative factors are not reflected in the probability. Real-time variables such as injuries and lineups are only partially captured right before game time.
· The live track record sample is still small (settled N=93) — calibration sharpens as the sample grows.
· Closing odds are not yet collected → CLV and edge are on hold.
· Qualitative factors are not reflected in the probability. Real-time variables such as injuries and lineups are only partially captured right before game time.
Frequently asked questions
Does SealedSlate's model beat the betting market?
No — and we publish that openly. Across a 1,350-game backtest the value-bet ROI was -6.77% and the model's log-loss (0.9771) was slightly worse than the closing market's (0.9497). SealedSlate's value is a tamper-proof, transparently published track record, not an edge over the market.
How is prediction accuracy measured?
With proper scoring rules applied to every settled prediction — Brier score and log-loss — plus a calibration table that checks whether, for example, 60%-confidence picks actually win about 60% of the time. Every figure is computed from the sealed ledger, not self-reported.
Can a past prediction be edited after the result is known?
No. Each prediction is sealed with a SHA-256 hash before kickoff and settled from official final results, so entries cannot be changed after the fact. The complete record is public on the Track Record page.
Every prediction is sealed with SHA-256 before the game — check the verdicts yourself on the Track Record. This service is for informational purposes only and does not solicit any specific wager.