Jul 26, 2025●13 reads●MIT License

Evaluation of Stock Market Prediction Techniques

100 days of code
Accuracy
AI Research
AI Showcase
ARIMA
Data Analytics
Data Science
Deep Learning
DJIA
Dow Jones Industrial Average
EDA
Exploratory Data Analysis
Feature Engineering
Financial Data
FinBERT
Forecasting Models
forecasting stock market
Git Bash
GitHub Repo
Hugging Face
Intent Engineering
Kaggle
LSTM
Machine Learning Models
ML Projects
Multimodal AI
NLTK
OpenSource AI
Predictive Analysis
Python
Quantitative Research
RedditNews
Sentiment Analysis
SHAP
stock
stock market
stock market forecast
stock market prediction
Stock Prediction
TCN
Technical Indicators
Time Series Forecasting
Vader

Abhishek

📊 Evaluation of Stock Market Prediction Techniques

This project compares the predictive performance of ARIMA, LSTM, and Temporal Convolutional Network (TCN) models on the Dow Jones Industrial Average (DJIA). A multimodal dataset combining technical indicators, sentiment scores, and intent signals from news and Reddit was engineered to forecast closing prices across three future horizons: t+1, t+3, and t+7.

📄 Read the Full Report (PDF)

📦 Dataset & Features

Dataset Size: ~1,900 trading days
Sources: Kaggle DJIA, financial news, Reddit discussions
Features:
- Technical: Log Returns, Volatility (Log10), Percentage Change
- Sentiment: VADER, FinBERT, Smart Sentiment
- Intent: Buying, Selling, Fear, Uncertainty, Urgency, etc.
Normalization: All sentiment and intent features scaled to [-1, 1] ranges

📂 Dataset on Hugging Face

⚙️ Modeling Approach

Prediction Targets:
- t+1
- t+3
- t+7
Train/Test Split: Chronological 80/20 split
Evaluation Metrics: R², RMSE, MAE
Pipeline:
- Feature engineering → Normalization
- Sequence generation (60 for LSTM, 30 for TCN)
- Model training → Testing → Residual analysis

🔍 Model Results (Train & Test)

ARIMA

Horizon	Train R²	Train RMSE	Test R²	Test RMSE
t+1	0.9921	238.87	0.8655	228.79
t+3	0.9877	297.54	0.7433	316.38
t+7	0.9800	380.06	0.5662	413.11

ARIMA excelled in short-term forecasts but declined sharply for longer horizons due to lack of external signal integration.

LSTM (Simple)

Horizon	Train R²	Train RMSE	Test R²	Test RMSE
t+1	0.9881	294.14	0.7606	319.58
t+3	0.9863	315.44	0.2669	559.01
t+7	0.9855	323.81	0.4156	501.07

Simple LSTM generalized well for t+1 but struggled beyond that, underfitting in volatile regimes.

LSTM (Stacked)

Horizon	Train R²	Train RMSE	Test R²	Test RMSE
t+1	0.9826	355.83	0.3267	535.94
t+3	0.9858	320.75	0.3985	506.37
t+7	0.9700	466.63	0.4252	496.95

While overfitting at t+1, stacked LSTM showed improved stability at longer horizons, especially t+7.

Temporal Convolutional Network (TCN)

Horizon	Train R²	Train RMSE	Test R²	Test RMSE
t+1	0.9953	186.12	0.8767	225.27
t+3	0.9910	257.10	0.7584	316.20
t+7	0.9831	351.86	0.5483	434.50

TCN consistently outperformed all other models at t+1 and t+3, showing superior temporal generalization and faster trend adaptation.

📊 SHAP Analysis (TCN)

t+1: Smart Reddit Sentiment, Greed Intent dominated
t+3: Reddit Prediction Intent, News Urgency Intent had highest weight
t+7: Aggregated sentiment took over, technical signals ranked low

🧠 Key Takeaways

Multimodal features (especially Reddit and news intent) boost predictive power.
TCN dominates short-term forecasting with minimal lag and high generalization.
ARIMA remains a solid short-term benchmark.
Stacked LSTM shows resilience in volatile mid-horizon targets.

🚀 Future Work

Integrate high-frequency trading and macroeconomic data
Explore hybrid ensembles (ARIMA + TCN + LSTM)
Add attention-based forecasting models like Informer/Transformer
Deploy as a live dashboard using Streamlit + Hugging Face APIs

📝 Citation

Malaviya, A. (2025). Evaluation of Stock Market Prediction Techniques. Dublin City University.
Report (PDF)