Predicting missing values for various financial datasets through machine learning for consistency and better model predictive power for a US-based Quantitative Hedge Fund
Client: Large US Quantitative Hedge Fund
Objective
To help a leading US-based quantitative systematic hedge fund improve data quality and enhance fundamental investment research by using Machine Learning techniques to impute missing values in financial statements.
CRISIL's Solution
- The client provided large data extract from a single market vendor for nearly 9,000 US companies
- CRISIL rapidly set up a team of Machine Learning and Accounting experts
- The team carefully assessed missing data and client requirements and proposed the MICE approach to impute missing values
- Used Expectation Maximization algorithm with statistical techniques like auto-regression and PMM to arrive at final results
- The model was validated using different sub-samples of data, using in-sample and out-sample techniques, and the validation of output against actual reported results using small sample of companies for key line items
- Model performance was measured using Mean Absolute Deviation and RMSE for benchmarking against using previous year values, pmm as imputation methodology
Client Impact
- CRISIL's MICE approach outperformed results from other vendors
- Enhanced quality data improved the Hedge Fund’s fundamental analysis of companies across equities/credit portfolios
Questions