Quantitative research · 2026
Point-in-Time Equity Drawdown Risk
Can a model rank the risk of a severe 15% drawdown over the next 40 trading days using only information available at each prediction date?
Rank drawdown risk without looking into the future
Rapid price gains sometimes end in severe drawdowns, but not every stock is equally exposed. The task was to rank that risk using only data available on the prediction date, while avoiding survivorship bias.
Build and freeze a point-in-time research pipeline
- Reconstructed historical index membership and corporate actions.
- Aligned prices and SEC fundamentals to when they were actually available.
- Used walk-forward evaluation so every prediction came from earlier training data.
- Froze the model before transferring it unchanged to the S&P 400.
The risk ordering held out of time
On 91,299 S&P 500 observations, the highest-risk decile had a 26.28% event rate versus 7.44% in the lowest-risk decile. The frozen model then transferred to 85,587 S&P 400 observations.
Replication summary
- S&P 500 ROC AUC
- 0.671
- Risk-decile ratio
- 3.53×
- S&P 400 ROC AUC
- 0.654
- Positive annual rank IC
- 7 / 7 years
[PASS] Frozen-model transfer criteria met
Drawdown risk was rankable; reliable alpha was not established
The model consistently separated lower- from higher-risk stocks, but the study did not establish reliable alpha. I also added monitoring and 84 automated tests so the pipeline can be rerun and checked.