Quantitative Trading: A First Look With QuantConnect

Can machine learning predict financial markets? This repo accompanies a 3-part video series where I test three generations of ML — from a 1970 linear model to a 2024 foundation model — on the same backtesting platform.

📺 Click here to watch the full playlist on YouTube!

---

Quick Navigation

Part 1: Classical ML (Ridge Regression)

Part 2: Deep Learning (Temporal CNN)

Part 3: Foundation Models (Amazon Chronos)

Note: When opening in Colab, you'll see a "This notebook was not authored by Google" warning — click Run anyway to proceed.

---

Part 1: Classical ML (Ridge Regression)

▶️ Watch the Part 1 Video

The first video covers ridge regression — a classical linear model from 1970 — applied to inverse volatility weighting on 12 futures contracts.

The strategy:

• Trade 12 futures (indices, energy, grains)
• Predict next-week volatility using ridge regression
• Allocate inversely: less volatile contracts get more capital
• Rebalance weekly

The result: Sharpe 0.212, Alpha -0.062. The model tracks the market with extra drawdown. It doesn't generate alpha.

Why it matters: Understanding why a simple model fails sets up everything that follows. Linear models can't capture the nonlinear patterns in financial data.

---

Part 2: Deep Learning (Temporal CNN)

▶️ Watch the Part 2 Video

The second video covers temporal convolutional networks — detecting patterns across multiple timescales in price data.

The strategy:

• Trade top 3 QQQ holdings (AAPL, MSFT, NVDA)
• 15-day OHLCV features → 3-class prediction (up/down/stationary)
• Temporal split: long/mid/short-term signals with separate learned weights
• Weekly retraining, confidence-weighted position sizing

The result: Sharpe 0.649, Alpha +0.093. The CNN found genuine signal with lower drawdown (26% vs 55%) and positive alpha — beating the market with less risk.

Why it matters: Nonlinear pattern detection works — but PSR 21.9% means we'd need a longer backtest to confirm it isn't noise. The forensic analysis in the notebook explores this limitation.

---

Part 3: Foundation Models (Amazon Chronos)

▶️ Watch the Part 3 Video

The third video covers zero-shot time series forecasting — using a pre-trained transformer that has never seen financial data.

The strategy:

• Trade top 5 stocks by dollar volume (AAPL, MSFT, NVDA, GOOGL, AMZN)
• Use Amazon Chronos to forecast 63-day price trajectories
• Sharpe-optimal portfolio weights via SciPy SLSQP
• Quarterly rebalancing

The result: Base model: Sharpe 0.727, Alpha +0.040. Fine-tuned (3 gradient steps): Sharpe 0.846, Alpha +0.076. Both beat ridge regression with zero training effort.

Why it matters: Foundation models transfer temporal patterns learned from weather, retail, and energy data to finance. But high beta (1.1) means most returns come from market exposure, and PSR 34.6% still isn't statistically significant. The notebook includes caveats on tokenization lookahead bias and trading frictions.

Note: The end-to-end pipeline cell requires a T4 GPU in Colab. Go to Runtime → Change runtime type → T4 GPU.

---

Local Setup

The notebooks use plotly for interactive charts. Everything else is narrative + embedded QuantConnect code.

git clone https://github.com/ababber/pyhou-02-17-2026.git
cd pyhou-02-17-2026
python -m venv .venv source .venv/bin/activate  # On Windows: .venv\Scripts\activate
pip install numpy scikit-learn plotly jupyter tensorflow pandas matplotlib
jupyter notebook

Requirements:

• Python 3.9+
• numpy, scikit-learn, plotly, jupyter
• tensorflow, pandas, matplotlib (Part 2)
• torch, chronos-forecasting (Part 3)

Note: Google Colab has most dependencies pre-installed. For Part 3, run !pip install chronos-forecasting in Colab. The end-to-end pipeline requires a T4 GPU runtime.

---

QuantConnect Setup

The actual trading strategies run on QuantConnect, a cloud-based algorithmic trading platform.

To run the strategies yourself:

• Create a free QuantConnect account at quantconnect.com
• Navigate to Algorithm Lab → Create New Algorithm
• Copy the algorithm code from the notebook (each technique has its full source)
• Run a backtest

Note: QuantConnect provides free backtesting with delayed data. The strategies in this series use futures contracts, which require a QuantConnect subscription for live trading but are free for backtesting.

| Part | QC Tier Required | |------|------------------| | 1 | Free | | 2 | Free | | 3 | Research¹ (fine-tuning only) |

¹ Part 3 requires a QuantConnect Researcher plan ($60/mo or $600/yr) for fine-tuning exercises. Backtesting the pre-trained model works on free tier.

For detailed QuantConnect setup, see pyhou-live-presentation-material/QC.md.

---

Origin

This material is based on a live presentation I gave at PyHou (Houston Python Meetup) on February 17, 2026.

• Original Meetup Event
• PyHou Meetup Page

---

References

Primary source:

• Pik, J., Chan, E. P., Broad, J., Sun, P., & Singh, V. (2025). Hands-On AI Trading with Python, QuantConnect, and AWS. Wiley. ISBN 978-1394268436.

Model origins:

• Hoerl, A. E. & Kennard, R. W. (1970). "Ridge Regression: Biased Estimation for Nonorthogonal Problems." Technometrics, 12(1), 55–67.
• LeCun, Y. et al. (1989). "Backpropagation Applied to Handwritten Zip Code Recognition." Neural Computation, 1(4), 541–551.
• Ansari, A. F. et al. (2024). "Chronos: Learning the Language of Time Series." arXiv:2403.07815.

Further reading:

• López de Prado, M. (2018). Advances in Financial Machine Learning. Wiley.
• Jansen, S. (2020). Machine Learning for Algorithmic Trading. Packt.

---

Disclaimer

This content is for educational purposes only. Nothing here is financial advice. Quantitative trading involves substantial risk of loss. Past performance does not guarantee future results. The strategies shown here lost money relative to a simple buy-and-hold benchmark.

---

License

MIT — see LICENSE for details.