Why some ideas stay in research
Some ideas make immediate hockey sense and still fail to improve the forecast. We look at whether the idea is plausible, whether it works on the narrow problem it was designed for, and whether it improves the complete simulation once the other inputs are present. Better player estimates, for example, do not always lead to better team-outcome forecasts.
When a useful signal adds little to the full model
Suppose an injury-onset model does a good job of identifying which players are more likely to miss time. That would be useful on its own, but adding an injury penalty to every future team path may still make the index forecast worse. Ordinary injury risk may already be reflected in historical team rates or market inputs, so another penalty can count the same risk twice.
A lineup model can also identify who played while still missing what happens next: who absorbs the available minutes, how effective that player is, and which situations he enters. Unless that transition is modeled well, cleaner lineup data may do little for the final distribution of team outcomes.
Research decisions as of September 4, 2026
These are summaries of existing research decisions, not a claim that every experiment has been retested against today’s exact source contract.
| Idea | Current decision | What would justify another look |
|---|---|---|
| Trade-deadline shocks | Disabled; the corrected test did not establish value. | Point-in-time inputs and a new passing comparison. |
| Future injury incidence | Not promoted into an unconditional injury penalty. | Improved joint forecast value without double counting baseline injury risk. |
| Shot-style and sequence features | Not promoted; limited benefit and incomplete source audit. | Better source evidence and held-out distribution improvements. |
| Direct historical lineup membership | Not adopted; tested forecasts worsened. | Better replacement and minute-allocation modeling. |
| Richer team context / odds trends | Retained as research, not default additions. | Consistent incremental gains on the relevant horizons. |
How this list relates to the current model
The composite reference currently includes xG, RAPM, and schedule-aware injury-return behavior. An earlier injury-only challenger did not beat the reference, so we still need a fresh attribution test to learn what the injury-return layer contributes inside the current model.
When we would revisit an idea
We keep an idea open when there is a specific reason the next test could turn out differently: better historical data, a more realistic replacement-player model, a corrected evaluation setup, or a clearer view of which forecast horizon should benefit. Hockey intuition is useful for forming the hypothesis, but the comparison still has to show an improvement.
For each revisit, define the proposed change before looking at the holdout, preserve the existing reference, and compare on the same information policy. Report both a layer-on/layer-off result and, when appropriate, how it interacts with other layers. A gain from the combined stack does not identify which ingredient caused it.
Some ideas should simply remain excluded until the necessary source exists. Deadline behavior that needs reliable historical cap or roster information is an example: filling missing values with plausible guesses would make a backtest look more complete while weakening its evidential value.
What we record from each test
State the question, frozen baseline, information cutoff, comparison method, relevant horizons, uncertainty, and decision. Report negative findings alongside positive ones. A feature earns promotion through evidence; a scenario control being available is not evidence of promotion.
Based on the internal current model state and September 4 model review. No new experiment or model promotion was performed for this article.
Independent educational research. Examples are hypothetical unless stated otherwise. Read current exchange rules and fund documents for product terms.
Read the full model overview →