Financial Markets Impact Models

The modeling results presented here indicate that - at least for Gold and the S&P 500 - the occurrence of earthquakes is a poor predictor of the direction of financial markets. Of primary importance is the measure known as accuracy: it measures how well a model performs when compared to a set of actual outcomes - values unknown to the model when it was devised. The model score is a measure of the internal efficacy of the model and is included here for completeness. The best result obtained was 9.21% accuracy; this was using an XGBoost model for Gold prices after a 6.7 earthquake. From there things get worse: Random Forest Regressor models perform significantly less effectively; and Linear and Logistic Regression models give especially poor results. Another interesting observation is that when the threshold is decreased from Magnitude 6.7+ to 5.5+ (resulting in an approximately twelvefold increase in the number of incidents used in the model), the accuracy decreased significantly. A final point is that the XGBoost models using the Mag 5.5+ dataset - running in a maxed-out ml.p3.16xlarge AWS Sagemaker notebook instance - ran approximately 12 days of CPU (yes!) for the S&P 500 model and 10 days for the Gold model.

NOTE: These results essentially confirm what numerous other sources have indicated - that indeed an earthquake typically has virtually no impact on the Financial Markets; see for example here, here and here. The one exception of note is the incident that occurred in Japan in March 2011, which saw the Nikkei 225 drop some 23% following the incident; but this may well be mitigated by two factors: one is the subsequent tsunami and the attendant nuclear reactor disaster, and the other is fact that the 93% of Nikkei 225 Futures traders were bullish just days before the earthquake hit - a notoriously bearish indicator.

Financial Instrument   Gold   S&P 500
Incident Magnitude   6.7+ 5.5+   6.7+ 5.5+
Incident Count   1445 23510   1870 28350
Model Type        
Linear Regression Model Score   0.009 0.003   0.011 0.001
RMSE   5.996 5.604   4.270 4.317
R-square   -0.012 0.001   -0.012 1.414
Logistic Regression Model Score   0.004 0.006   0.003 0.005
Accuracy   0.00% 0.59%   0.21% 0.31%
Random Forest Regressor Model Score   0.912 0.989   0.895 0.986
Accuracy   6.14% 1.33%   4.42% 1.30%
XGBoost Model Score   0.144 0.485   0.181 0.449
Accuracy   9.21% 6.05%   7.28% 5.12%