Case study

Abnormal Move Prediction in S&P 500 Stocks

Will a stock make an abnormal move tomorrow? A binary classifier over 5 years of daily price and volume data for 505 S&P 500 stocks, built for CSE 5367 (Pattern Recognition) at UT Arlington.

GitHub

Test AUC (Random Forest): 0.723
Test F1 at a 2% threshold: 0.307
Daily OHLCV rows: 619,040
No-lookahead features: 12

How it works

1
Load
619,040 daily OHLCV rows for 505 stocks, 2013–2018
2
Engineer features
12 per stock-day (returns, volatility, price range, gap, true range), past data only
3
Label
Next-day return beyond 2% or 1.5% in either direction
4
Split by time
Train, validation and test in date order, no shuffling
5
Train and tune
Five models, hyperparameters chosen on validation F1
6
Evaluate
All models scored on the same held-out test period

Results

Test set at the 2% threshold (12,455 abnormal and 106,397 normal stock-days)
Model	Test F1	AUC	Precision	Recall
Logistic regression, L1	0.303	0.716	0.226	0.460
Logistic regression, L2	0.303	0.716	0.226	0.456
Linear SVM	0.300	0.715	0.229	0.435
Random Forest	0.307	0.723	0.221	0.502
Logistic regression, L2 + PCA	0.301	0.715	0.226	0.451

What I found

Random Forest was best at both thresholds, which suggests nonlinear feature interactions matter.
L1 and L2 regularization gave nearly identical results. L1 zeroed 2 of the 12 features without hurting performance, so those carried little independent signal.
PCA at 95% variance cut 12 features to 9 components and slightly hurt performance: the dropped components held a small amount of real signal.
Volatility and price-range features were consistently the strongest predictors.

Limitations

Validation F1 was higher than test F1 for every model. The test year (2017) was a calmer market than the validation year (2016), which is a real cost of a time-based split.
All stochastic steps use a fixed random seed, so the numbers above reproduce from the repo.

Figures

ROC curves for the five models at the 2% threshold — ROC curves at the 2% threshold.

Random Forest feature importances at the 2% threshold — Feature importances at the 2% threshold.

Built with

PythonPandasNumPyScikit-learnTime-Series ValidationFeature Engineering