Portfolio Return
−1.17%
₹−11,653 · 2-day window
Best Stock
HDFCBANK
+0.26% · only green stock
Worst Stock
M&M
−3.09% · ₹−4,570 loss
Best Model (MAPE)
ARIMA
0.97% avg backtest MAPE
Best Direction
Prophet
55.2% avg directional acc.
Stock Universe — 5 NSE Stocks
| Stock | Sector | Buy Price | Day 2 Close | 2-Day Return |
| HDFCBANK |
Banking |
₹745.03 | ₹746.93 |
+0.26% |
| INFY |
IT |
₹1,122.67 | ₹1,116.91 |
−0.51% |
| SUNPHARMA |
Pharma |
₹1,861.14 | ₹1,817.18 |
−2.36% |
| HINDUNILVR |
FMCG |
₹2,252.21 | ₹2,239.52 |
−0.56% |
| M&M |
Auto |
₹3,217.50 | ₹3,118.16 |
−3.09% |
Models Used in This Capstone
| Model | Type | Purpose | Avg MAPE | Avg Dir.Acc | Role |
| ARIMA |
Statistical |
Price level forecasting |
0.97% |
49.5% |
Primary |
| Prophet |
Additive |
Direction signals + seasonality |
6.37% |
55.2% |
Direction |
| LSTM |
Deep Learning |
Non-linear pattern capture |
1.88% |
46.6% |
Supporting |
| XGBoost |
Gradient Boost |
Lag-feature ML forecasting |
~2.1% |
~52% |
ML Baseline |
| HMM |
Probabilistic |
Regime detection — Bull / Bear |
— | — |
Context |
| GARCH |
Volatility |
Risk estimation for Task 5 |
— | — |
Risk |
| Markowitz |
Optimization |
Efficient frontier & max Sharpe |
— | — |
Allocation |
Backtest Performance — Jul to Dec 2025 · 130 trading days
Directional Accuracy % — Higher is Better
RMSE in Rupees — Lower is Better
Full Comparison Table
| Stock |
ARIMA MAPE | Prophet MAPE | LSTM MAPE | XGB MAPE |
ARIMA Dir | Prophet Dir | LSTM Dir | XGB Dir |
Winner |
| HDFCBANK |
0.57% | 3.70% | 1.08% | ~1.2% |
51.6% | 54.0% | 47.6% | ~53% |
ARIMA |
| INFY |
1.01% | 9.37% | 1.57% | ~1.8% |
48.4% | 57.3% | 49.2% | ~50% |
ARIMA |
| SUNPHARMA |
1.41% | 9.17% | 1.81% | ~2.0% |
51.6% | 62.1% | 52.4% | ~54% |
Prophet |
| HINDUNILVR |
0.83% | 3.19% | 1.72% | ~1.5% |
43.6% | 53.2% | 50.8% | ~52% |
ARIMA |
| M&M |
1.02% | 6.43% | 3.23% | ~2.4% |
52.4% | 49.2% | 33.1% | ~54% |
XGBoost |
Model Verdict
ARIMA wins price accuracy (0.97% avg MAPE) — best for price-level targets and allocation weights.
Prophet wins directional accuracy (55.2%) — best for buy/sell signals.
XGBoost is a strong all-rounder (~52.6% directional) — best ML baseline.
LSTM improves with multivariate technical indicators.
Portfolio strategy: ARIMA prices + Prophet direction + GARCH risk weights.
2-Day Trading Forecasts — May 12 & 13, 2026
All Model Forecasts vs Actual Prices
| Stock | Last (May 11) |
ARIMA May 12 | Prophet May 12 | LSTM May 12 | XGB May 12 |
Actual May 12 | Actual May 13 |
| HDFCBANK | ₹780.85 |
₹778.72 | ₹848.02 | ₹783.15 | ~₹779 |
₹745.56 | ₹746.93 |
| INFY | ₹1,179.20 |
₹1,179.20 | ₹1,168.29 | ₹1,177.35 | ~₹1,175 |
₹1,119.04 | ₹1,116.91 |
| SUNPHARMA | ₹1,847.90 |
₹1,842.98 | ₹1,714.04 | ₹1,744.07 | ~₹1,840 |
₹1,858.27 | ₹1,817.18 |
| HINDUNILVR | ₹2,287.70 |
₹2,289.10 | ₹2,269.78 | ₹2,289.30 | ~₹2,285 |
₹2,252.71 | ₹2,239.52 |
| M&M | ₹3,330.40 |
₹3,335.67 | ₹3,482.24 | ₹3,145.43 | ~₹3,300 |
₹3,211.92 | ₹3,118.16 |
Forecast Error — May 13 Actual vs Predicted (%)
Key Finding: Train–Test Distribution Shift
All models predicted prices above actual on May 13, 2026.
ARIMA backtest MAPE was 0.97% but live error jumped to
~4.1%. A broad NSE decline on May 13 created a
market-wide distribution shift that no univariate time series model
could detect from price history alone.
Fix: Add macro features (Nifty 50, FII/DII flows, India VIX)
as exogenous regressors in ARIMAX / Prophet to detect market-wide sentiment shifts.
Final Allocation — Task 5 Blend
Strategy Comparison — Weights
Markowitz Efficient Frontier
Portfolio Metrics Comparison
| Portfolio | Ann. Return | Ann. Volatility | Sharpe Ratio | Max Weight | Notes |
| Max Sharpe (Markowitz) |
~18% | ~20% |
~0.65 | HDFCBANK |
Optimal risk-adjusted |
| Min Variance (Markowitz) |
~12% | ~16% |
~0.38 | HINDUNILVR |
Lowest possible risk |
| Equal Weight |
~15% | ~22% |
~0.44 | Equal |
Naive benchmark |
Portfolio Insight
Our Task 5 blend sits close to the efficient frontier — between Equal Weight and Max Sharpe.
GARCH-based inverse volatility weighting correctly overweighted HDFCBANK (lowest risk) and
underweighted M&M (highest risk). This was validated by actual returns: HDFCBANK +0.26%, M&M −3.09%.
Total Invested
₹9,72,709
5 stocks · 2-day window
Final Value (Day 2)
₹9,61,056
At market close May 13
Total P&L
−₹11,653
−1.17% overall
Live ARIMA Error vs Backtest — May 13
Directional Accuracy — Live May 12 & 13
| Stock | Actual Direction | ARIMA | Prophet | LSTM | XGBoost |
| HDFCBANK |
UP ↑ |
↓ ✗ | ↓ ✗ |
↓ ✗ | ↓ ✗ |
| INFY |
DOWN ↓ |
→ ✗ | ↓ ✓ |
↑ ✗ | ↑ ✗ |
| SUNPHARMA |
DOWN ↓ |
↑ ✗ | ↓ ✓ |
↓ ✓ | ↑ ✗ |
| HINDUNILVR |
DOWN ↓ |
↑ ✗ | ↓ ✓ |
↑ ✗ | ↑ ✗ |
| M&M |
DOWN ↓ |
↑ ✗ | ↑ ✗ |
↓ ✓ | ↑ ✗ |
Key Reflection
Prophet outperformed all models on live directional accuracy (60%) —
correctly predicting DOWN for INFY, SUNPHARMA, and HINDUNILVR.
The broad market decline on May 13 was a regime shift that
HMM would flag as a Bear transition if monitored daily.
HDFCBANK validated our low-volatility allocation strategy —
only green stock, only stock below GARCH threshold.
M&M and SUNPHARMA — highest GARCH volatility — were
worst performers exactly as the risk model predicted.