Setup
We pulled 12,400 head-to-head pairs with at least 8 prior meetings and measured how well each historical result predicted the next one, weighted by recency.
What we found
- Last 1–3 meetings: strongest signal, decaying by ~30% per step
- Last 4–5 meetings: moderate residual signal
- 6+ meetings back: no statistically meaningful contribution once form and league context are controlled for
How the model uses it
H2H is included as a bounded feature capped at the last five meetings, with an exponential recency weight. Older meetings contribute to the archive display but not to the current probability distribution.