Forecasting Short-Term NIFTY Returns

A Comparative Study of Naive, HMM, LSTM and Hybrid HMM-LSTM Models

---

Project Overview

Can deep learning predict short-term stock index returns?

This project investigates one of the most common assumptions in quantitative finance:

Can a deep learning model outperform a simple statistical baseline when forecasting the next 25-minute return of the NIFTY 50 Index?

To answer this question, four forecasting models were evaluated using a strict chronological train-validation-test split:

• Naive Baseline
• Hidden Markov Model (HMM)
• Long Short-Term Memory Network (LSTM)
• Hybrid HMM + LSTM

Rather than optimizing until a positive result appeared, the goal was to objectively evaluate whether increasingly sophisticated models genuinely improved forecasting performance.

---

Dataset

• NIFTY 50 Index
• Two years of historical 1-minute OHLCV data
• Resampled to 5-minute candles
• Approximately 37,000 observations after feature engineering

---

Feature Engineering

The forecasting models were trained using technical and volatility-based features including:

• Log Returns
• Rolling Volatility (5, 10 and 20 bars)
• ATR Ratio
• EMA Distance (10,20,50)
• EMA Spread
• RSI(14)

For the HMM, only market-state variables were used:

• Log Return
• Volatility
• ATR Ratio

---

Forecasting Pipeline

1-Minute Data │ ▼ 5-Minute Resampling │ ▼ Feature Engineering │ ▼ Predict 25-Minute Return │ ▼ ──────────────────────────────── Naive Baseline │ LSTM │ Gaussian HMM │ Hybrid HMM + LSTM ──────────────────────────────── │ ▼ Out-of-Sample Evaluation

---

Models Compared

1. Naive Baseline

Assumes the next return is approximately equal to the current return.

This provides the minimum benchmark every forecasting model should outperform.

---

2. Gaussian Hidden Markov Model

Learns latent market regimes directly from market behaviour using:

• Returns
• Volatility
• ATR

Instead of predicting returns directly, the HMM predicts hidden market states.

For forecasting, each hidden state predicts the average future return historically observed in that regime.

---

3. LSTM

A recurrent neural network trained using:

• 10-candle lookback window
• 32 hidden units
• Dropout Regularization
• Adam Optimizer
• Early Stopping

---

4. Hybrid HMM + LSTM

The hidden market regime discovered by the HMM is one-hot encoded and appended to the LSTM feature vector.

This tests whether latent market-state information improves neural network forecasting.

---

Evaluation Metrics

Each model was evaluated using:

• Mean Absolute Error (MAE)
• Root Mean Squared Error (RMSE)
• R² Score
• Directional Accuracy
• Sharpe Ratio
• Maximum Drawdown
• Final Equity Curve

A simple long-short signal strategy with transaction costs was also simulated to evaluate economic usefulness.

---

Results

The experiment produced an unexpected result.

While the Hybrid HMM + LSTM consistently improved upon the standalone LSTM, none of the machine learning models outperformed the naive baseline out-of-sample.

This highlights an important lesson in quantitative finance:

Increasing model complexity does not necessarily improve predictive performance.

---

Key Takeaways

• Chronological validation is essential for financial machine learning.
• Naive benchmarks are surprisingly difficult to beat.
• HMM-derived market regimes improved neural network performance but not enough to outperform a simple baseline.
• Short-horizon return prediction remains an extremely challenging problem.

---

Repository Structure

Forecasting-Short-Term-NIFTY-Returns/

│ ├── data/ │ ├── results/ │ ├── model_comparison.csv │ ├── feature_correlation.png │ ├── hmm_regimes.png │ ├── residuals_grid.png │ ├── training_curves.png │ ├── predicted_vs_actual_grid.png │ ├── model_performance_summary.png │ └── comparison_equity_curves.png │ ├── main.py │ ├── requirements.txt │ └── README.md

---

Future Work

This project intentionally concludes with a negative result.

Rather than continuing to optimize return prediction, future work focuses on:

• Market regime classification
• Probabilistic forecasting
• Multi-horizon market state prediction
• Regime-aware decision support systems

Those ideas are implemented in the follow-up project:

NIFTY Market State Engine

---

Technologies Used

• Python
• TensorFlow / Keras
• hmmlearn
• NumPy
• Pandas
• Scikit-learn
• Matplotlib
• Seaborn

---

Author

Anand Rai

Quantitative Finance • Machine Learning • Systematic Trading

If you found this project useful, consider starring the repository.