Stocker – NSE Stock Price Prediction

June 2025 - Ongoing (Final Year Project)
End-to-end Streamlit web app for NSE stock forecasting - currently under active development with multiple ML models, PostgreSQL metrics storage, and horizon-optimized predictions spanning 1-week to 5-year horizons using RMSE, MAPE, F1, and Cohen's Kappa evaluation.

Problem Statement

• Traditional stock prediction fails across varying time horizons - trees excel short-term momentum, linears survive long-term drift
• Investors need symbol+horizon specific model recommendations not generic ensembles
• Nifty 50 lacks comprehensive multi-metric benchmarks combining price error + directional accuracy
• Single-model apps ignore performance degradation patterns critical for practical deployment

System Architecture

Modular pipeline flows from Streamlit UI through ML backend to data and persistence layers. User selects Nifty 50 stock triggering cached yfinance fetch, automated feature engineering with 20+ technical indicators, parallel training of 8 regression models, multi-horizon metric computation, PostgreSQL persistence, averaged-rank model selection, and autoregressive forecasting. Caching dramatically accelerates retraining while database enables cross-session leaderboards and performance evolution tracking.

Data Pipeline

• Live NSE equities loaded from official CSV (2000+ valid .NS symbols)
• 5-year daily OHLCV via yfinance(period='5y', interval='1d')
• Extracts Date/Close columns, validates minimum 100 points
• Chronological 80/20 train-test split preserves time-series integrity
• Recent 10-day data preview confirms quality before training

Feature Engineering

Temporal, momentum, and technical features engineered avoiding leakage:

• Temporal: dayofweek, month, dayofyear, weekofyear seasonality
• Momentum: Close lags 1-5 days + 7-day rolling mean/standard deviation
• RSI(14): 14-day gain/loss ratio for overbought/oversold detection
• MACD: EMA12-EMA26 difference + 9-period signal line
• Bollinger Bands: SMA20 ± 2×std20 with bandwidth for volatility squeezes All features shifted once before target creation (next Close) yielding ~1200 clean samples.

Modeling Approach

Eight benchmark regression models trained with fixed hyperparameters for fair comparison:

• Tree Ensembles: XGBoost(100), RandomForest(100), AdaBoost(100)
• Instance-Based: SVR(RBF), KNN(k=5)
• Parametric: LinearRegression, DecisionTree(depth=6)
• Neural: MLP(64→32 hidden, 500 iterations) XGBoost gracefully falls back if unavailable. Models fit chronologically ordered training data predicting next-day closing price.

Multi-Horizon Strategy

Realistic investor timeframes tested via sliding windows on chronological test set:

• 1-week: last 5 trading days (momentum dominant)
• 1-month: last 22 days (trend persistence)
• 1-year: last 252 days (mean reversion begins) Metrics computed per window reveal characteristic degradation patterns across horizons.

Model Selection Methodology

Averaged ranking across four metrics per horizon ensures robust selection:

1. Individual ranks: min(RMSE, MAPE), max(F1, Kappa)
2. Mean rank calculation across all available metrics
3. Lowest average rank wins per horizon Example results: Random Forest dominates 1-week, XGBoost 1-month, SVM survives 1-year Bar charts highlight best models in gold with 3-decimal precision annotations.

Database & Persistence Layer

PostgreSQL stores comprehensive model_results table capturing:

• 16 metrics per stock/index/model/horizon combination
• Auto-deletes prior evaluations per symbol ensuring fresh results
• Safe float conversion handles NaN/infinity values
• Pandas read_sql() queries generate instant leaderboards Enables cross-session performance comparison and model evolution tracking.

Current Status: Active development by team of 4 - targeting LSTM integration, backtesting framework, hyperparameter optimization, and paper publication by project completion.