Tennis Glicko

Model Domain & Architectural Scope

Architectural Limits: Why Pre-Match Models Invalidate in Live, Play-by-Play Environments

Pre-match win probabilities model static prior distributions (P(A wins | I_pre)). The moment a match starts, intra-match stochastic dynamics require point-level Markov Chain state transitions (P(A wins | S_t)).

Machine Learning & Stochastic Modeling~7 min read

1. The Theoretical Boundary of Pre-Match Features

The XGBoost model powering Tennis Glicko is trained on historical pre-match feature matrices I_pre, comprising surface-adjusted Glicko-2 ratings, Elo ratings, head-to-head history, rest days, and tournament tier dynamics.

Mathematically, the output is a prior probability distribution P(Y=1 | I_pre). It answers a specific question: Given historical performance up to time t_0, what is the probability that Player A wins the match?

This formulation explicitly excludes intra-match state variables $S_t$ (e.g., service hold rates in current set, physical distress, weather/wind shifts during play, break point conversions).

2. Point-by-Point State Transitions: Markov Chains vs Static Gradient Boosting

Live, in-play tennis modeling requires dynamic state-space tracking. A tennis match can be modeled as a discrete-time Markov Chain where the state S_t = (g_A, g_B, s_A, s_B, server) tracks games, sets, and points.

In a live environment, the win probability updates conditional on point outcomes:

P(A wins | S_t) = p_serve * P(A wins | S_{t+1}^win) + (1 - p_serve) * P(A wins | S_{t+1}^loss)

Because static pre-match XGBoost architectures cannot dynamically update state transitions during a match, attempting to apply pre-match model outputs or VOPO deltas to live markets constitutes an out-of-domain statistical error.

3. Summary of Model Application Boundaries

USE CASESTATUSRATIONALE
Pre-Match Line EvaluationVALIDFeatures strictly match pre-game information state I_pre.
Historical BacktestingVALIDEvaluates out-of-sample Brier Score against closing lines.
Live In-Play ExecutionINVALIDATEDLacks Markov point-by-point dynamic state updating.
Micro-Betting / Next PointINVALIDATEDModel does not estimate intra-game serve/return probabilities.

Return to Learn Index

Browse all research guides, methodologies, and technical documentation.

Learn Index →