Context-Relative Action Scoring

Measuring what a player achieved relative to the distribution of what was historically achievable in similar situations. Instead of giving a player their raw EPV delta (+0.5 for a through ball), find all historically similar possession sequences, compute the distribution of outcomes in those situations, and report where this player's action falls in that distribution (e.g., 85th percentile). This fixes the positional-reward-access-bias: a defender's break-even play in a context where 90% of players got negative outcomes is 90th percentile performance, not a zero.

The pipeline has three stages: (1) Sequence encoding: compress each possession or event sequence into a latent vector using an LSTM autoencoder — the encoder feeds in raw event sequences (not just geometry, but event types, directions, pressure), squeezes through a bottleneck, and the decoder tries to recreate the original sequence; the bottleneck representation captures the "meaning" of the possession. (2) Similarity matching: for each event, find K nearest neighbor sequences in latent space (or cluster via K-means if compute is limited). (3) Percentile scoring: compute the player's action value delta, compare it to the distribution of historical deltas in matched sequences, and report a percentile rank scaled to [-1, +1].

Important: the autoencoder should capture more than geometry — include event types, pass types (through ball vs. cross vs. carry), pressure state, and sequence ordering. The LSTM structure preserves temporal information that flat geometric approaches miss. If K-nearest-neighbors is too slow (millions of sequences × millions of comparisons), fall back to K-means clustering, but accept the fidelity loss.

1. Using only geometric similarity: Two possessions that look identical on a heat map can be completely different in tempo, event sequence, and pressure context. The autoencoder must encode more than coordinates.
2. Using z-scores instead of percentile ranks: Z-scores produce extreme outliers in rare contexts. Percentile ranks bounded to [-1, +1] are more robust.
3. Clustering too coarsely: K-means with too few clusters loses the fidelity of individual situation matching. Use as many clusters as compute allows, or go full K-nearest-neighbors if feasible.