Efficient Market Hypothesis in the Intraday Time Horizon

Abstract

This thesis employs machine learning models to predict the S&P500 ETF’s (SPY) last half-hour’s return. In order to make their predictions, these models discover and utilize novel momentum and reversal trends whose existence is inconsistent with the Efficient Market Hypothesis. Their impressive out-of-sample performance support the notion that trading behavior earlier in the day can be indicative of how the market will act in its final half-hour of trading. The main machine learning models applied are the LASSO model and the decision tree model. Econometric analysis of the regressions they utilize reveal that the discovered trends are highly statistically significant. All price and volume data for the SPY available at the time of collection from the Trade and Quote’s millisecond database (9/10/2003 to 3/13/2018) was used to create these models’ independent variables, which represent intuitive financial indicators for the behavior of the last half-hour.