METR

Market Exposure Timing Research

A personal research project exploring if machine learning can improve simple trading signals.

Status: Finished  •  License: MIT

---

What is this project?

"Can a machine learning model beat random market entries using just price and volume data?"

METR is a personal experiment I built to see if I could use machine learning to filter out bad trades from a basic momentum strategy. I wanted to test this rigorously, so I compared my model's performance against 10,000 random Monte Carlo simulations to make sure any success wasn't just luck.

---

How it works

• The Core Idea: I start with a basic momentum signal (e.g., buy if the price has been going up). Then, I use an XGBoost classifier to look at other indicators and decide whether to actually take the trade or pass on it.
• Labeling: I used a "Triple Barrier" method to label my data. It sets a profit target, a stop loss, and a time limit (like 5 days) for each trade.
• Features: I used Polars to build technical indicators, including some fractional differentiation to make the data stationary while keeping its history.
• Validation: I checked the results using standard metrics like Sharpe ratio and ran Monte Carlo simulations to verify the stats.

---

Results: GoldBees

The most interesting result came from the Gold ETF (GoldBees). The model was actually able to turn a losing basic momentum signal into a profitable one on unseen data from 2024–2025.

Performance Metrics (Out-of-Sample 2024–2025)

Metric	Result
Return	+17.92%
Sharpe Ratio	1.48
Max Drawdown	7.99%
Win Rate	66.1%

I ran a few statistical tests (like a Monte Carlo simulation with 10k runs), and the results for Gold were statistically significant (p-value ~0.01). However, the model didn't find any real edge for Nifty 50 or USD/INR, which makes sense given how efficient and managed those markets are.

---

Project Structure

Click to view

METR/
├── data/                    # Raw and processed data
├── docs/                    # My notes and methodology
├── src/
│   ├── fetch_data.py        # Fetches data from yfinance
│   ├── feature_eng.py       # Indicator logic using Polars
│   ├── triple_barrier.py    # Trade labeling logic
│   ├── train_trade_filter.py # The XGBoost model
│   ├── backtest_engine.py   # Code to calculate returns
│   └── monte_carlo_audit.py # Random simulations for testing
├── config.yaml              # Project settings
└── README.md                

---

Documentation

If you want to read more about what I learned and how I built it:

1. Documentation & Methodology: The main write-up of my findings.
2. Observations: My daily logs while working on the project.

References

• Triple Barrier Labelling Algorithm by William Santos – Implementation guidance for volatility-adaptive labeling.
• Fractional Differencing Financial Data by The Quant Trading Room – Theoretical foundation for memory preservation.

---

_{Built by Mohak Kapoor}