SHF-ML: Machine Forecast Disagreement

Executive Summary

SHF-ML is an empirical asset pricing study that measures forecast disagreement across information sets. Drawing on the Machine Forecast Disagreement (MFD) literature and the Jensen-Kelly-Pedersen (JKP) characteristic dataset, it simulates heterogeneous investor beliefs with two Random Forest forecasters: one on accounting features, one on market features.

Problem & Constraints

The question: if two investors see different categories of information, do they produce systematically different forecasts for next month's excess returns? It runs on Seawulf and NVwulf HPC with 10-year rolling windows and leakage-safe preprocessing. Code lives on university lab systems (not publicly distributed). Reference: JKP Factors Database.

Methodology

• Used JKP Monthly Dataset (Jensen-Kelly-Pedersen Global Factor Database, 2015-2025)
• Partitioned 153 stock characteristics into accounting-only and market-only subsets
• Trained separate Random Forest models targeting ret_exc_lead1m
• Computed forecast disagreement as the difference between accounting and market predictions
• Deployed reproducible pipeline on Seawulf/NVwulf with partitioned batch jobs

Results & Metrics

Tech Stack

Python, Random Forest, scikit-learn, JKP dataset, SeaWulf HPC, NVwulf cloud

Future Work

Complete paper replication (portfolio tests, asset-pricing analysis), expand to 100-investor random feature subsets per MFD paper, publish methodology documentation.

Links

• JKP Factors Database
• Replication Crisis Repository
• JKP Documentation (PDF)