Stock Price Forecasting using LSTM Neural Networks (NIFTY-50 Dataset)

Stock-price prediction is one of the most picked — and most poorly executed — final-year ML topics. Typical builds shuffle time-series data before splitting (leaking the future into training), report a single accuracy-style number with no baseline, and collapse the moment an examiner asks why the model should beat "tomorrow looks like today". This project takes the disciplined route: real NIFTY-50 daily OHLCV data from the NSE, strictly chronological splits, scalers fit on training data only, and a persistence baseline running on the same evaluation folds so the model's true edge is explicit. A sequence-to-one LSTM learns from 15-day windows of open, high, low, close and volume to regress the next day's closing level, evaluated with RMSE, MAE and MAPE under walk-forward retraining. A live web demo makes the whole pipeline tangible: the price explorer charts the real series, and a pure-JavaScript LSTM trains in the browser while loss curves and forecast-vs-actual plots update live.

1. Daily OHLCV for NIFTY-50 (^NSEI) is fetched from NSE archives via yfinance — 249 trading days (27 Sep 2024 – 26 Sep 2025) in the reference build.
2. Each training sample becomes a 15-day window of 5 features (open, high, low, close, volume); the target is the next trading day's close.
3. The series is split chronologically 80/20 with no shuffling, and min-max scalers are fit on the training portion only.
4. A Keras model — LSTM(32) → Dropout(0.2) → Dense(1, linear) — trains with Adam (learning rate 0.005) on mean squared error, with early stopping on a validation slice.
5. Walk-forward evaluation retrains on an expanding window and scores each held-out fold with RMSE, MAE and MAPE.
6. The persistence baseline (predict yesterday's close) is scored on the identical folds, so any edge — or lack of one — is stated plainly in the report.
7. The web demo embeds the series and trains a smaller pure-JS LSTM live in the browser, plotting training/validation loss and the one-step-ahead forecast against actual closes.

Tech stack:

• Python 3, TensorFlow/Keras
• yfinance (NSE data fetch), pandas, NumPy
• scikit-learn (scaling, metrics)
• Matplotlib (evaluation plots)
• Jupyter Notebook (training & evaluation)
• HTML/CSS/JavaScript (in-browser LSTM demo)

Dataset & model details

• Dataset: NIFTY-50 index (^NSEI) daily OHLCV from the National Stock Exchange of India via Yahoo Finance. Reference build: 249 trading days, 27 Sep 2024 – 26 Sep 2025, 5 feature columns per day (Open, High, Low, Close, Volume).
• Task: Univariate time-series regression — input = 15-day window × 5 features, output = next trading day's closing level in index points.
• Model: Sequence-to-one LSTM: LSTM(32, tanh/sigmoid gates) → Dropout(0.2) → Dense(1, linear); approximately 5,000 parameters; Adam optimizer (lr 0.005), MSE loss, early stopping.
• Metrics: RMSE, MAE and MAPE on chronological held-out folds under walk-forward retraining, reported alongside the persistence baseline. Design target: MAPE around 1% of the index level. No metric is claimed as measured — the build report records the figures achieved by the training run executed for the order.

• [Real NIFTY-50 data pipeline] Daily OHLCV for the NIFTY-50 index (^NSEI) pulled from NSE archives via yfinance into a cleaned, documented CSV — no synthetic or hand-made price series anywhere.
• [Leakage-free sequence preparation] 15-day sliding windows with a strictly chronological 80/20 split; min-max scalers are fit on the training portion only, exactly as a deployed forecaster would require.
• [LSTM forecaster (Keras)] Single-layer LSTM with 32 units, dropout 0.2 and a linear Dense(1) output — roughly 5,000 parameters, trained with Adam on MSE with early stopping.
• [Walk-forward evaluation] Expanding-window retraining with RMSE, MAE and MAPE scored per fold, so performance is measured the way a real forecasting workflow measures it.
• [Persistence baseline] "Tomorrow equals today" runs on the same folds — the honest first bar every price model must clear, and a strong viva talking point when it is hard to beat.
• [In-browser live training demo] A pure-JavaScript LSTM (forward pass, backpropagation through time, Adam, dropout) trains on the embedded real series with live loss curves and a forecast-vs-actual overlay.
• [Price explorer dashboard] Interactive chart of the full series with 20/50-day moving averages, volume overlay and OHLC hover detail.
• [Viva-ready documentation] Report, PPT and Q&A covering LSTM gates, why shuffling time series is wrong, and how to interpret RMSE/MAPE honestly.

• Python data-fetch and preprocessing scripts (yfinance → cleaned NIFTY-50 CSV)
• Training and walk-forward evaluation Jupyter notebook
• Trained Keras model file with fitted scalers
• Live web demo: price explorer, architecture views, in-browser LSTM training with loss curves and forecast overlay
• Evaluation plots: loss curves, forecast-vs-actual, baseline comparison
• Project report PDF (background, LSTM theory, methodology, honest results discussion)
• PPT presentation for final review
• Viva Q&A preparation document (LSTM gates, leakage, baselines, metric interpretation)

• Daily index prices behave close to a random walk — a well-tuned LSTM may not beat the persistence baseline on every fold; the report shows both numbers honestly.
• Single-day horizon only: the model predicts the next close, not multi-day trajectories, which compound error quickly.
• Fitted on one index and one market regime — elections, crashes or policy shocks degrade any model trained on calm-period data.
• The ~1% MAPE figure is a design target, never a measured claim; the achieved metrics are recorded after the build run for your order.
• yfinance/NSE data revisions or missing sessions must be re-validated on each fresh data pull; the pipeline documents the check.
• This is a forecasting study, not a trading system: no backtester, no transaction costs, no position sizing — nothing here implies profitability.