Simulating Investor Disagreement with Machine Learning

The insight from MFD literature

The Machine Forecast Disagreement paper argues that investor disagreement is hard to observe directly. So instead of surveying beliefs, represent investors as different ML models fed different information subsets. Forecast dispersion becomes the proxy for disagreement.

Our simplified framework

Instead of 100 investors with random feature subsets, SHF-ML uses two:

• Investor A - accounting information only
• Investor B - market information only

Both forecast next month's excess return (ret_exc_lead1m) on the JKP Monthly Dataset using Random Forests deployed on Seawulf HPC.

Disagreement measure

For each stock-month: Prediction_A(i,t) - Prediction_M(i,t)

High disagreement may predict lower future returns when short-selling is difficult - testing the MFD hypothesis on a reproducible HPC framework.

Takeaway

This is a working replication framework, not yet a full paper replication. The Project Champion award recognized building reproducible infrastructure before claiming empirical results.