Time Series Analysis Capstone Project: Portfolio Allocation

A Python-based portfolio optimization system using LSTM neural networks, ARIMA baseline models, and GARCH volatility weighting for Indian stock market analysis.

📋 Project Overview

This project implements a strict train/test validation framework for forecasting 9 Indian NSE stocks (RELIANCE, TCS, HDFCBANK, SUNPHARMA, ITC, BHARTIARTL, LTTS, ASIANPAINT, ICICIBANK) using:

1. LSTM (Stacked Long Short-Term Memory) - Deep learning time series forecaster
2. ARIMA(5,1,0) - Baseline statistical model for comparison
3. GARCH(1,1) - Volatility forecasting for portfolio weighting
4. ADF Test - Stationarity diagnostics
5. STL Decomposition - Seasonal-Trend analysis

Capital Limit: ₹10,00,000 (10 lakh rupees)

---

📁 File Structure

market_data.py (Data Pipeline)

Central module for all data operations. Implements the strict train/test split as required by the academic rubric.

Key Components:

• STOCKS - List of 9 NSE tickers
• TARGET_STOCKS - Subset of 6 stocks for ADF analysis
• TRAIN_START = '2021-01-01'
• TRAIN_END = '2025-06-30'
• TEST_START = '2025-07-01'
• TEST_END = '2025-12-31'

Functions:

• fetch_data(start, end) - Downloads Yahoo Finance close prices with auto-adjustment and forward-fill handling
• split_train_test(data) - Splits data into training (Jan 2021 – Jun 2025) and test (Jul–Dec 2025) periods
• get_latest_price(stock_name, fallback_series) - Real-time price with 3-tier fallback strategy

---

Model.py (Main Orchestrator)

Executable portfolio engine. Run this to validate, forecast, and allocate capital.

Workflow:

1. Fetch Data - Download 5+ years of historical prices
2. Validate Models - Compare LSTM vs ARIMA on the explicit test period
3. Calculate Metrics - MAPE, RMSE, Directional Accuracy for each model
4. Forecast - Next 2 trading days using LSTM
5. Weight Portfolio - Inverse-volatility allocation using GARCH
6. Print Allocation Table - Stock, Current Price, Forecasted Price

Key Functions:

• train_lstm(series) - Train stacked LSTM (50→50 neurons, 2 Dropout layers)
• forecast_lstm(model, scaler, series, steps) - Rolling forecast for n steps ahead
• forecast_arima(series, steps, order=(5,1,0)) - ARIMA baseline
• calculate_metrics(actual, predicted) - MAPE%, RMSE, Direction Accuracy%
• validate_models(data) - Print comparison table
• get_volatility_weight(series) - GARCH(1,1) inverse-volatility weighting
• run_system(data) - Execute full portfolio pipeline

Architecture (LSTM):

Input(60×1) 
  → LSTM(50, return_seq=True) 
  → Dropout(0.2) 
  → LSTM(50, return_seq=False) 
  → Dropout(0.2) 
  → Dense(1 output)

Training: 5 epochs, batch_size=32, Adam optimizer, MSE loss

---

otherModels.py (Preprocessing & Diagnostics)

Time series analysis and stationarity testing.

Functions:

• check_stationarity(stock_name, series) - ADF test (p-value ≤ 0.05 = stationary)
• plot_decomposition(stock_name, series) - STL decomposition with 252-day seasonal period
• run_other_models() - Execute ADF for all target stocks, STL plot for BHARTIARTL.NS

Output:

• Terminal: ADF statistics and p-values
• File: bhartiartl_ns_stl.png - Seasonal-Trend-Residual plot

---

🚀 How to Run

Prerequisites

Ensure your Python environment is configured:

pip install tensorflow pandas numpy scikit-learn yfinance statsmodels arch

Execute Main Pipeline

python Model.py

Output:

1. Data fetch confirmation
2. VALIDATION REPORT table: | Stock | LSTM_MAPE | LSTM_RMSE | LSTM_Dir% | ARIMA_MAPE | ARIMA_RMSE | ARIMA_Dir% | Winner |
3. Win summary (LSTM vs ARIMA count)
4. Per-stock analysis (current price, forecast, volatility weight)
5. FINAL PORTFOLIO ALLOCATION table: | Stock | Final Weight (%) | Number of Shares | Current Price | Forecasted Price |

Run Time Series Analysis

python otherModels.py

Output:

• ADF test results for all 6 target stocks
• bhartiartl_ns_stl.png plot (if 252+ observations available)

---

📊 Key Parameters

Parameter	Value	Purpose
SEQ_LENGTH	60 days	LSTM lookback window
CAPITAL	₹10,00,000	Portfolio investment amount
LSTM epochs	5	Training iterations
LSTM batch_size	32	Samples per update
GARCH(p,q)	(1,1)	Volatility model order
ARIMA order	(5,1,0)	Baseline AR-I-MA configuration
STL period	252 days	Trading days per year

---

📈 Output Explained

Validation Table Columns

• LSTM_MAPE - Mean Absolute Percentage Error (%)
• LSTM_RMSE - Root Mean Squared Error (₹)
• LSTM_Dir% - Directional accuracy (% correct up/down predictions)
• ARIMA_MAPE, ARIMA_RMSE, ARIMA_Dir% - Same metrics for baseline
• Winner - Model with lower MAPE on the test period

Portfolio Allocation Columns

• Stock - NSE ticker name
• Final Weight (%) - Inverse-volatility normalized portfolio weight
• Number of Shares - Integer shares sized using ₹10,00,000 capital limit
• Current Price - Real-time price (3-tier fallback: intraday → fast_info → historical)
• Forecasted Price - LSTM prediction for next trading day

Allocation Strategy

1. Train LSTM + GARCH on full history
2. Forecast next 2 trading days
3. Calculate volatility (GARCH)
4. Weight inversely: weight = 1/volatility / Σ(1/volatility_all)
5. Allocate capital proportionally

---

🔧 Data Handling

• Missing Data: Forward-fill then backward-fill (ffill().bfill())
• Price Adjustment: Yahoo Finance auto_adjust=True (splits/dividends included)
• Series Validation: Drops NaN, checks minimum length thresholds
• Scaler: MinMaxScaler (0,1) for LSTM training only

---

✅ Rubric Compliance

• Task 2 & 3 (Train/Test Split): ✓ Strict Jan 2021 – Jun 2025 train; Jul–Dec 2025 test
• Task 4 (Preprocessing): ✓ ADF stationarity tests, STL decomposition
• Task 5 (Portfolio Sizing): ✓ GARCH weighting, ₹10L capital allocation
• Task 6 & 8 (Baseline & Metrics): ✓ ARIMA(5,1,0), MAPE/RMSE/Direction Accuracy
• Task 2 (Data Pipeline): ✓ Centralized market_data.py with ffill/bfill

---

📝 Notes

• The test period (Jul–Dec 2025) is explicitly hard-coded in market_data.py for reproducibility
• LSTM is retrained on each run; results vary slightly due to random initialization
• GARCH may warn if volatility is very low; falls back to std deviation
• STL decomposition requires ≥252 observations (skipped otherwise)
• Real-time price fetches may fail in offline environments; uses historical fallback

---

📦 Dependencies

tensorflow>=2.21.0
pandas>=3.0.3
numpy>=2.4.5
scikit-learn>=1.8.0
yfinance>=1.3.0
statsmodels>=0.14.6
arch>=8.0.0

---

📞 Troubleshooting

Issue	Solution
ImportError: No module named 'tensorflow'	pip install tensorflow
yfinance: No data found	Check ticker format (use .NS for NSE)
GARCH fitting error	Reduce returns smoothing or skip stock
STL insufficient data	Need ≥252 trading days (about 1 year)
Real-time price fetch timeout	Script falls back to historical close

---

🎓 Academic Context

This project demonstrates:

• Time series forecasting with deep learning (LSTM)
• Statistical baseline comparison (ARIMA)
• Volatility modeling (GARCH)
• Portfolio optimization principles
• Proper train/test validation methodology
• Handling real market data with YFinance API

Capstone Focus: Rigorous train/test separation with explicit dates, side-by-side model evaluation, and production-grade error handling.