Projects: Sera
==============

Time-series machine learning for trading · Inter-IIT Tech Meet internal
selection (a past Inter-IIT problem) · Sep – Oct 2026

Framing
-------

I framed the brief as a causal three-class prediction problem (Hold / Long /
Short) on second-level data for two anonymized instruments with masked features.
Each decision uses only past information.

Features
--------

I researched and proposed these engineered features:

- a Kalman-filter smoothed-price slope
- a CUSUM-style regime state
- sine/cosine encoding of trading time

Pipeline
--------

The pipeline trains an XGBoost classifier in memory-bounded batches over ~9.3M
sequences. A cost-aware backtester scores the results.