Volatility-Adaptive Position Sizing — Implementation Guide

Instruments: Crude Oil futures (CL) and 10-Year T-Note futures (ZN) Goal: Forecast next-day realized volatility, convert each forecast into a dollar-risk-targeted position size, and backtest whether the sizing rules kept risk disciplined.

---

What this repo contains

• data/fetch_data.py — fetches CL and ZN history from yfinance and computes log returns + realized volatility.
• models/garch_model.py — rolling GARCH(1,1) forecasts.
• models/features.py — time-series features for the ML model.
• models/ml_model.py — LightGBM regression with walk-forward forecasting.
• risk/position_sizing.py — volatility-targeted sizing engine with a 10-lot cap.
• risk/fair_value.py — simple cost-of-carry fair-value flag module.
• backtest/backtest_engine.py — builds the simulation frame for both models.
• backtest/metrics.py — realized-risk metrics, breach rate, RMSE, and drawdown utilities.
• dashboard/app.py — Streamlit dashboard skeleton for CL and ZN.
• requirements.txt — dependencies for running the project.

---

Quick start

cd /Users/manangarg/Desktop/Finance_Project/Volatility-Adaptive-Position-Sizing
python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install --upgrade pip
python3 -m pip install -r requirements.txt
python3 -m streamlit run dashboard/app.py

If you prefer to validate the pipeline manually first, use:

from data.fetch_data import load_or_fetch, compute_realized_vol
from backtest.backtest_engine import build_backtest_frame

df = load_or_fetch('CL=F', period='3y')
df = compute_realized_vol(df)
bt = build_backtest_frame(df, 'CL')
print(bt[['realized_vol', 'garch_vol', 'ml_vol', 'garch_size', 'ml_size']].tail())

This manual check confirms the data fetcher, feature pipeline, forecasts, and sizing outputs before you build the dashboard.

---

Project summary

This project is a risk-management system, not a directional signal.

• GARCH is the baseline volatility forecast.
• LightGBM is the ML forecast using lagged realized vol, range-based volatility, volume, and calendar features.
• The sizing rule targets a fixed dollar risk per trade ($500) and caps positions at 10 lots.
• The backtest measures whether realized risk stayed close to the target, not whether price predictions were right.

---

Data

Source: yfinance for CL=F and ZN=F daily OHLCV.

The pipeline computes:

• log_return = daily log return of close prices
• realized_vol = 10-day rolling annualized standard deviation of log returns

This realized volatility is the ground truth used for one-step-ahead forecasting.

---

Models

GARCH(1,1)

• Built with arch
• Uses a rolling fit and walk-forward forecasting to avoid lookahead bias
• Forecasts the next day’s volatility from past return shocks and variance

LightGBM ML forecast

• Uses features computed from historical data available at each time point
• Walk-forward training is used instead of random splits
• The model is trained on lagged volatility, Parkinson range, volume momentum, and calendar signals

---

Position sizing

The core sizing rule is:

position_size = target_risk_dollars / (forecast_daily_vol × price × contract_multiplier)

Where:

• target_risk_dollars = $500
• contract_multiplier = $1,000 for both CL and ZN
• forecast_daily_vol = daily volatility forecast as a decimal
• price = current futures price
• maximum allowed lots = 10

This keeps the maximum notional risk aligned with the target and maps to a realistic TT SIM cap.

---

Backtest description

For each day in the backtest:

1. compute past realized volatility and time-series features
2. produce a GARCH forecast and an ML forecast using only prior data
3. compute the lot size for each model
4. calculate the next-day realized dollar risk for that size
5. compare realized risk against the $500 target

This is a risk-discipline backtest, not a P&L direction backtest.

---

What "verify the model pipeline manually" means

It means running the core code directly, not only the dashboard. A manual check ensures:

• historical data loads correctly
• realized volatility is computed correctly
• GARCH forecasts are generated in a walk-forward way
• ML features and forecast outputs line up
• position size recommendations are produced for both models

That gives confidence in the pipeline before adding plots or UI.

---

Initial results from the current implementation

These are the first validation metrics from a 3-year backtest run.

• CL=F:

  • usable backtest days: 482
  • GARCH RMSE vs realized vol: 0.4338
  • ML RMSE vs realized vol: 0.0756
  • GARCH mean absolute realized-risk deviation: 1,290.67
  • ML mean absolute realized-risk deviation: 500.00
  • both models currently produce conservative sizing in the initial validation

• ZN=F:

  • usable backtest days: 483
  • GARCH RMSE vs realized vol: 0.0459
  • ML RMSE vs realized vol: 0.0021
  • GARCH mean absolute realized-risk deviation: 6,251.998
  • ML mean absolute realized-risk deviation: 500.00
  • initial sizing remains conservative, with the latest ML recommendation at 0 lots

Note: these are preliminary numbers from the implementation as of this run. They should be updated after improving the ML walk-forward training and the size calculation logic.

---

Next steps

• improve the ML feature set and walk-forward split
• compute full backtest metrics including drawdown and tolerance breach rate
• add the fair-value flag to the dashboard panels
• make the dashboard charts and model comparison tables interactive
• document the final findings and interview walkthrough clearly

---

Notes

• keep CL and ZN separate in all analysis
• keep the 10-lot cap explicit
• the fair-value module is a secondary signal, not the main sizing engine
• any resume bullet should be backed by the actual final backtest run