08-12-2026, 08:58 PM
Machine learning can help classify regimes, rank instruments, or combine features, but it does not remove the need for clean data and strict validation.
Start with a simple baseline: a rules-based or linear model using a small feature set. Then compare any complex model against that baseline on the same point-in-time data and out-of-sample windows.
Avoid leakage from target construction, normalisation across future rows, and random train/test splits that mix time periods. Use time-ordered splits. Log feature availability, model version, prediction timestamp, and confidence.
A model should be useful after costs, uncertainty, and operational complexity—not just accurate in a notebook. What baseline do you use before testing a more complex model?
Start with a simple baseline: a rules-based or linear model using a small feature set. Then compare any complex model against that baseline on the same point-in-time data and out-of-sample windows.
Avoid leakage from target construction, normalisation across future rows, and random train/test splits that mix time periods. Use time-ordered splits. Log feature availability, model version, prediction timestamp, and confidence.
A model should be useful after costs, uncertainty, and operational complexity—not just accurate in a notebook. What baseline do you use before testing a more complex model?
Building small AI tools and automations in Bengaluru. Notes on what ships, what fails, and what I'm learning along the way.