Jingyi (Joy) Zhang

Computational Projects

Elo Ratings as Constant-Step SGD

GitHub
Problem
How does the Elo update factor trade off stationary rating noise against adaptation when latent skill changes?
Approach
Controlled Bradley-Terry simulations and a chronological ATP forecasting study interpret Elo as constant-step stochastic gradient descent.
Result / status
On 39,247 completed ATP matches from 2010-2025, validation selected a surface-aware Elo model. Its 2022-2025 test log loss was 0.6195 versus 0.6225 for standard Elo; the project reports no statistically resolved improvement over betting-market probabilities.
Three-panel chronological ATP study showing the Surface Elo validation grid, held-out model log losses, and test-period calibration.
Chronological ATP study: validation-selected Surface Elo, locked 2022–2025 test results, and calibration.
  • Stochastic approximation
  • Bradley-Terry models
  • Walk-forward validation
  • Python