[Paper Review] A Study on Stock Forecasting Using Deep Learning and Statistical Models

Himanshu Gupta, Aditya Jaiswal|arXiv (Cornell University)|Feb 8, 2024

Stock Market Forecasting MethodsDecision Sciences3 citations

TL;DR

This study evaluates deep learning and statistical models—ARIMA, RNN, LSTM, CNN, and WaveNet-based Full CNN—on S&P 500 data for stock price forecasting. Using mean absolute error (MAE) as the primary metric, LSTM with a 30-day window achieved the lowest MAE of 1.18, outperforming all other models, including ARIMA (MAE 2.8) and Full CNN (MAE 7.98).

ABSTRACT

Predicting a fast and accurate model for stock price forecasting is been a challenging task and this is an active area of research where it is yet to be found which is the best way to forecast the stock price. Machine learning, deep learning and statistical analysis techniques are used here to get the accurate result so the investors can see the future trend and maximize the return of investment in stock trading. This paper will review many deep learning algorithms for stock price forecasting. We use a record of s&p 500 index data for training and testing. The survey motive is to check various deep learning and statistical model techniques for stock price forecasting that are Moving Averages, ARIMA which are statistical techniques and LSTM, RNN, CNN, and FULL CNN which are deep learning models. It will discuss various models, including the Auto regression integration moving average model, the Recurrent neural network model, the long short-term model which is the type of RNN used for long dependency for data, the convolutional neural network model, and the full convolutional neural network model, in terms of error calculation or percentage of accuracy that how much it is accurate which measures by the function like Root mean square error, mean absolute error, mean squared error. The model can be used to predict the stock price by checking the low MAE value as lower the MAE value the difference between the predicting and the actual value will be less and this model will predict the price more accurately than other models.

Motivation & Objective

• To compare the performance of deep learning models (LSTM, RNN, CNN, WaveNet) and statistical models (ARIMA, moving averages) in forecasting stock prices.
• To identify the most accurate model for S&P 500 index prediction using quantitative error metrics like MAE, RMSE, and MSE.
• To evaluate the impact of window size and data preprocessing on model accuracy in time series stock forecasting.
• To determine whether deep learning models consistently outperform traditional statistical models in volatile, non-stationary financial time series.

Proposed method

• Trained and tested ARIMA, RNN, LSTM, CNN, and Full CNN models on historical S&P 500 data (open, high, low, close, volume).
• Used a sliding window approach with 20-day and 30-day window sizes for LSTM and CNN models to capture temporal dependencies.
• Implemented a WaveNet-inspired architecture in the Full CNN model with dilated convolutions (dilation rates 1, 2, 4, 8) to capture long-term patterns.
• Applied data preprocessing techniques including normalization and differencing to reduce non-stationarity in stock price time series.
• Evaluated model performance using MAE, RMSE, and MSE, with lower MAE indicating higher predictive accuracy.
• Used Python with NumPy, Keras, scikit-learn, and matplotlib for model development, training, and visualization.

Experimental results

Research questions

• RQ1 Which model—statistical (ARIMA) or deep learning (LSTM, RNN, CNN)—achieves the lowest mean absolute error (MAE) in S&P 500 stock price forecasting?
• RQ2 How does window size (20 vs. 30 days) affect the performance of LSTM and CNN models in predicting stock prices?
• RQ3 Does the WaveNet-based Full CNN model outperform a standard pre-processed CNN in capturing temporal patterns in stock data?
• RQ4 How do RNN variants (sequence-to-vector vs. sequence-to-sequence) compare in forecasting accuracy for financial time series?
• RQ5 To what extent do deep learning models reduce prediction error compared to naïve and moving average baselines?

Key findings

• LSTM with a 30-day window achieved the lowest MAE of 1.18, outperforming all other models tested, including ARIMA (MAE 2.8).
• The Full CNN model (MAE 7.98) significantly outperformed the pre-processed CNN (MAE 12.65), demonstrating the effectiveness of the WaveNet architecture.
• Sequence-to-sequence RNN achieved a lower MAE (4.19) than sequence-to-vector RNN (23.07), indicating better long-term dependency learning.
• ARIMA performed reasonably well with an MAE of 2.8 but was outperformed by all deep learning models, especially LSTM.
• The naïve model (using previous day’s price) served as a strong baseline, but all deep learning models improved upon it.
• LSTM’s ability to capture long-term dependencies and non-linear patterns in noisy, non-stationary stock data led to superior performance across metrics.