Stock market candlestick charts on trading monitors

IEEE Research Paper · Co-Author

Intraday Market Analysis and Forecasting with LSTM Networks

2025 5th International Conference on Advancement in Electronics & Communication Engineering (AECE) · November 2025

PythonLSTMTime SeriesPandas

This paper asks whether a model can forecast a stock’s daily high and low price using nothing but its own trading history — no news sentiment, no macroeconomic indicators, just the numbers the market has already produced.

The dataset was ten years of daily price data for a publicly traded company (2015–2025), pulled from Yahoo Finance. The values were normalized with Min-Max scaling, then reshaped into 1,500-timestep sequences — long historical windows the model could draw patterns from.

Line chart of a publicly traded company's daily high price from 2015 to 2026 — Fig. 2 — ten years of daily high-price history, the raw signal the model learns from.

The architecture uses LSTM (Long Short-Term Memory) networks, a type of recurrent neural network built to recognize patterns across sequences over time. Rather than one model predicting both the day’s high and low, two dedicated LSTM models were trained separately, each with the Adam optimizer and mean squared error loss.

System diagram: stock data flows through EDA, preprocessing, LSTM training, evaluation, and ten-day forecasting — The end-to-end pipeline — raw stock data through preprocessing, LSTM training, evaluation, and a 10-day forecast.

HIGH-PRICE MODEL

R² = 0.954

MAE = 34.01

RMSE = 44.26

LOW-PRICE MODEL

R² = 0.947

MAE = 35.26

RMSE = 46.41

Chart comparing actual vs predicted low prices, closely tracking each other — Fig. 7 — actual vs. predicted low prices on held-out data. The model tracks real price movement closely.

An R² above 0.94 on both models means the networks explain the large majority of the variance in next-day price movement using historical pattern alone. The paper was co-authored and published to IEEE Xplore through the AECE conference.

View on IEEE Xplore