This experiment evaluates the impact of stale model calibrations on tail risk estimates from the perspective of statistical backtesting.
Here, daily log stock returns are modeled as a t-distribution Levy process. Model parameters are estimated via maximum likelihood using exponentially decaying flexible probabilities.
The backtesting impact is then measured by how the p-value distribution of a binomial backtest shifts when using a stale calibration instead of a freshly updated one.
- presentation/ contains slides for the presentation itself
- src/ contains the Python implementation
- scripts/ contains scripts that demonstrate model usage and generate the presentation content
A written version of this presentation is available at https://ryanagibson.com/posts/stale-calibration-backtest/ and an informal ~10-minute version of this discussion is available at https://www.youtube.com/watch?v=pU5SRbmKAOg.