Deep Learning · Finance · Time Series

🧠 Stock Price Prediction
with LSTM

Forecasting weekly closing prices of Barclays PLC (BARC.L) using a stacked LSTM neural network trained on 2.5 years of historical market data.

🐍 Python 3.7+ 🔴 Keras / TensorFlow 📊 Pandas · NumPy 🔬 scikit-learn

Overview

📌 Project Summary

This project builds a Long Short-Term Memory (LSTM) neural network to forecast stock prices — one of the most challenging problems in financial machine learning. Unlike traditional models, LSTMs learn long-range temporal dependencies, making them well-suited for financial time series where past behaviour influences future price movements.

The model is trained on weekly OHLCV data for Barclays PLC (LON: BARC) from January 2020 to August 2022, covering the COVID-19 crash, the recovery rally, and the 2022 macro downturn — a highly varied test of the model's generalisation ability.

---

Data

📈 Closing Price History

Full BARC.L weekly close price history. The vertical dashed line marks the 80/20 train/test boundary.

Fig 1 — BARC.L weekly closing prices (Jan 2020 – Aug 2022), 137 data points total.

---

Model

🏗️ LSTM Architecture

A stacked two-layer LSTM followed by fully connected Dense layers for regression output.

Fig 2 — Model data flow: 60-step input window → LSTM×2 → Dense layers → single price prediction.

Layer	Type	Units	Notes
Input	—	60 × 1	60-week lookback window
LSTM 1	LSTM	50	return_sequences=True
LSTM 2	LSTM	50	return_sequences=False
Dense 1	Dense	12	Hidden layer
Dense 2	Dense	1	Price output

---

Preprocessing

📐 Data Preparation

Raw close prices are MinMax scaled to [0, 1] before training to stabilise gradient updates. The dataset is split into 80% training and 20% testing with no data leakage. Each training sample uses a 60-timestep sliding window to predict the next week's price.

Fig 3 — Scaled price data with 80% training (blue) and 20% test (orange) regions highlighted.

from sklearn.preprocessing import MinMaxScaler
scaler = MinMaxScaler(feature_range=(0, 1))
scaled_data = scaler.fit_transform(dataset)
training_data_len = math.ceil(len(dataset) * .8)  # → 110 weeks
# Build sliding windows: [t-60 ... t-1] → predict t
for i in range(60, len(train_data)):
x_train.append(train_data[i-60:i, 0])
y_train.append(train_data[i, 0])

---

Training

🔁 Model Training

Compiled with the Adam optimiser and Mean Squared Error loss. Trained for 2 epochs with a batch size of 2.

Fig 4 — Training vs validation MSE loss over epochs. Loss drops sharply from 0.1082 → 0.0053.

Epoch	Training Loss (MSE)
1	0.1082
2	0.0053

---

Results

🎯 Predictions vs Actual

Predicted prices (red dashed) overlaid against actual prices (green) across the held-out test set. The shaded region shows the prediction error band.