Dinesh Jinjala

12Personal research

Quant Research Platform

Author, private

My research platform for trading strategies. It covers 157K+ instruments on FastAPI and PostgreSQL, fetches market history on demand and computes trend, momentum and volatility features. I use it to test a volatility-contraction breakout strategy with an XGBoost entry model and a 1D-CNN sequence model, validated with purged k-fold splits and an embargo.

157K+
instruments searchable
Purged k-fold
with embargo, no look-ahead
XGBoost + 1D-CNN
entry and sequence models

Python / XGBoost / PyTorch / FastAPI / PostgreSQL

Problem

Backtests of trading ideas are easy to fool: overlapping time windows and look-ahead leaks make a strategy look better than it is. This platform tests a volatility-contraction breakout strategy with validation designed to rule out those leaks.

Approach

  • A FastAPI and PostgreSQL service can search 157K+ instruments and fetches OHLCV history from minute to monthly bars in chunks, upserting it into per-instrument tables.
  • A YAML-configured feature engine computes trend, momentum, volatility, price-action and market-structure features.
  • The strategy combines pattern detectors with an XGBoost entry model and a 1D-CNN sequence model, plus market regime and breadth gates and a backtest with a moving-average trailing exit.
  • Model validation uses purged k-fold time splits with an embargo, so no training sample overlaps the period it is tested on.
  • A fundamentals screener parses snapshots into normalized tables and builds point-in-time features.

Architecture

  1. Market data
  2. PostgreSQL store
  3. Feature engine
  4. XGBoost + 1D-CNN
  5. Purged k-fold
  6. Backtest & reports

Outcome

The platform runs the full loop from raw market data to validated backtests. I write up null results too, including the fundamentals and similarity experiments, instead of hiding them. It is a research tool, and I make no return claims from it.

What I learnedA null result you can trust is more useful than a backtest that only looks good.

Building something like this?

Tell me about it