How the Early-Warning System Works
U.S. · Can today’s market information warn of broad stress within the next 10 trading days?
01 Market Data
SPY, VIX and the Baa–Treasury spread describe equity, volatility and credit conditions.
View data details
The credit proxy is Moody’s Baa corporate yield minus the 10-Year U.S. Treasury yield. It is not an ICE BofA high-yield OAS series.
02 Market Stress Index
MSI equally weights equity, volatility and credit stress using past-only standardization.
Technical definition
Equity stress is the negative 21-trading-day SPY return; volatility stress is the VIX level; credit stress is the Baa–Treasury spread.
Each component is standardized using a 252-trading-day rolling window shifted by one observation. Standardized components receive equal weights.
Stress thresholds use information available at each point in time, rather than future full-sample information. Stress describes broad market conditions, rather than a simple future SPY decline.
03 17 Features
7 equity, 5 volatility and 5 credit features capture returns, volatility, momentum and relative stress.
Feature details
Equity · 7
SPY 1D Return · SPY 5D Return · SPY 20D Return · SPY 20D Volatility · SPY 60D Volatility · SPY Drawdown · SPY 20D Momentum
Volatility · 5
VIX Level · VIX 1D Change · VIX 5D Change · VIX 20D Momentum · VIX 252D Z-Score
Credit · 5
BAA–Treasury Spread Level · Credit 1D Change · Credit 5D Change · Credit 20D Momentum · Credit 252D Z-Score
Rate feature ablation
Four U.S. Treasury-rate and 10Y–2Y yield-curve features were tested as a 21-feature extension under the identical purged walk-forward OOS framework.
Logistic ROC-AUC declined from 0.922 to 0.916 and PR-AUC from 0.807 to 0.786. XGBoost ROC-AUC changed from 0.899 to 0.901 while PR-AUC declined from 0.779 to 0.773. No consistent incremental predictive value was found, so the final specification retains 17 features.
04 10-Day Forward Target
Target₁₀D(t) = 1 if at least one defined Stress Event occurs during t+1 through t+10; otherwise 0.
Target construction
The subsequent 10 trading days provide the training label. Features and event thresholds retain their point-in-time, past-only design.
Observations without a complete forward horizon remain unknown and are not converted to artificial negative labels.
05 Logistic Regression / XGBoost
Logistic Regression is the baseline; XGBoost models nonlinear relationships and produces the deployed score.
Model details
Historical scores are generated from out-of-sample predictions. The fitted XGBoost deployment model produces the latest score and SHAP drivers; Logistic Regression remains a benchmark.
Existing OOS findings show stronger discrimination and recall for Logistic Regression, while XGBoost has lower Brier error and fewer false alarms at the fixed 0.50 reference. These are different operating trade-offs.
Review actual model performance →06 Purged Walk-Forward Validation
Expanding training windows and a 10-day purge separate past training labels from unseen OOS tests.
Validation design
Random train/test splitting is not used. A 10-trading-observation gap removes training labels whose forward target windows overlap the beginning of the test period.
Evaluation uses ROC-AUC, PR-AUC, Brier Score, Precision, Recall, F1, False Alarm Rate and Missed Event Rate. Classification uses the fixed 0.50 raw-score reference (50 / 100 display scale), rather than a threshold selected post-hoc from OOS results.
07 Risk Intelligence
Model scores and SHAP support early warning without implying calibrated event probabilities or investment advice.
Limitations
SHAP identifies features that push the fitted XGBoost output higher or lower. Contributions are model-output effects, not probability-point changes, and do not establish causality.
Future performance can differ from historical OOS results. The system does not guarantee Stress Events or predict future SPY price direction.
Calibration diagnostics, particularly in intermediate and high-score regions, do not support interpreting raw outputs as literal event probabilities. Credit spreads are proxies for credit conditions.
Model outputs are decision-support signals, not investment recommendations. Scores are not necessarily calibrated event probabilities.