[論文レビュー] Predicting Stock Market Time-Series Data using CNN-LSTM Neural Network Model

A Aadhitya, R. Rajapriya+2 more|arXiv (Cornell University)|May 21, 2023

Stock Market Forecasting Methods被引用数 11

ひとこと要約

本論文は、株式市場の時系列データを予測するCNN-LSTMニューラルネットワークを提案し、1-Dの株式特徴をCNN用のテンソルへ変換、複数のデータセットで評価し、リアルタイムデータ展開を考慮する。

ABSTRACT

Stock market is often important as it represents the ownership claims on businesses. Without sufficient stocks, a company cannot perform well in finance. Predicting a stock market performance of a company is nearly hard because every time the prices of a company stock keeps changing and not constant. So, its complex to determine the stock data. But if the previous performance of a company in stock market is known, then we can track the data and provide predictions to stockholders in order to wisely take decisions on handling the stocks to a company. To handle this, many machine learning models have been invented but they didn't succeed due to many reasons like absence of advanced libraries, inaccuracy of model when made to train with real time data and much more. So, to track the patterns and the features of data, a CNN-LSTM Neural Network can be made. Recently, CNN is now used in Natural Language Processing (NLP) based applications, so by identifying the features from stock data and converting them into tensors, we can obtain the features and then send it to LSTM neural network to find the patterns and thereby predicting the stock market for given period of time. The accuracy of the CNN-LSTM NN model is found to be high even when allowed to train on real-time stock market data. This paper describes about the features of the custom CNN-LSTM model, experiments we made with the model (like training with stock market datasets, performance comparison with other models) and the end product we obtained at final stage.

研究の動機と目的

• Motivate and design a CNN-LSTM approach to predict stock market time-series data.
• Explore preprocessing strategies to convert 1-D stock data into CNN-friendly tensors.
• Evaluate model performance across diverse datasets and real-time data sources.
• Compare CNN-LSTM with other models and discuss deployment considerations.

提案手法

• Construct a custom CNN-LSTM architecture with 3 CNN layers (64,128,64) and kernel size 3, followed by 2 Bi-LSTM layers with 100 units each.
• Convert 1-D stock data (Close column emphasized) into [100,1] tensors using a 100-row moving window and feature-mean encoding.
• Train with Mean Squared Error loss, use Adam optimizer, and evaluate with MSE, MAE, and R2 metrics.
• Preprocess data from Alpha Vantage TIME_SERIES_DAILY full mode and Kaggle datasets, handle NULLs, and perform scaling/reshaping for TimeDistributed CNN.
• Save the model in HDF5, and provide Docker and Helm deployment for local and Kubernetes environments.

実験結果

リサーチクエスチョン

• RQ1 How well does a CNN-LSTM architecture capture features and temporal patterns in stock time-series data compared to other models?
• RQ2 What preprocessing strategies (tensor construction and windowing) improve CNN effectiveness on 1-D stock data?
• RQ3 How does the model perform across diverse stock markets (NYSE, NASDAQ, NIFTY) and real-time data sources?
• RQ4 What deployment considerations (Docker/Kubernetes) are practical for real-world usage of the CNN-LSTM stock predictor?

主な発見

Model / Dataset	Metric	Value (example)	Notes
CNN-LSTM (Our model)	MSE	0.035	Avg across datasets (Table IV)
LSTM (Our model)	MSE	0.045	Table IVで比較
XGBoost	MSE	0.047	Table IVで比較
NIFTY (SBIN - sample)	MSE	0.001	Table IIIから
NASDAQ (ACTG - sample)	MSE	0.1565	Table IIIから
NYSE (IBM - real)	MSE	0.0027	Table IIIから

• MSE scores around 0.001–0.035 across datasets, with MAE ~0.075 and R2 ~0.935 in some tests.
• Table II reports MSE 0.035, MAE 0.075, Variance 0.935370, R2 0.9353, Max Error 0.174930 for the tested setup.
• Table III shows dataset-specific MSE scores (e.g., NIFTY 0.001, NASDAQ 0.1565, IBM 0.0027, RELIANCE 0.0145).
• Table IV compares CNN-LSTM to custom models, with CNN-LSTM MSE 0.035 vs LSTM 0.045 and XGBoost 0.047.
• Table V positions CNN-LSTM against other models with MSE 0.035 (avg) vs various baselines (e.g., LBL-LSTM 0.017 train / 0.026 test, LSTM decentralized 0.0003).
• The paper demonstrates deployment via Docker images and Helm charts, enabling Kubernetes deployment and CI-based image publishing.