Hybrid Deep Learning Model for Multi- Asset Financial Market Prediction using LSTM- GRU Architecture

Hybrid Deep Learning Model for Multi- Asset Financial Market Prediction using LSTM- GRU Architecture

Prem Kumar Gupta

M. Tech, Computer Science and Engineering Department of Computer Science and Engineering Sarvepalli Radhakrishnan University, Bhopal, India

Dr. Varsha Namdeo

Professor, Department of Computer Science and Engineering Sarvepalli Radhakrishnan University, Bhopal, India

Abstract – Predicting financial market trends is kind of hard because time series data is always shifting around and behaves non linearly, which makes it unpredictable. In this work we put forward a Hybrid LSTMGRU Deep Learning setup, meant for forecasting prices of multiple assets, like shares, cryptocurrencies, foreign exchange, and even commodities. We gather past market information, then we clean it up and scale it with normalization and we also do feature engineering. For those features we lean on common technical signals such as RSI, MACD, SMA, and EMA. The hybrid method blends the long horizon memory strength of LSTM with the faster computational handling of GRU, so that the overall prediction quality is higher. After training, we judge the model using MAE, RMSE, MAPE, and R² Score, then we place its results next to both more traditional machine learning methods and other deep learning models. In the experiments, the hybrid design shows better forecasting accuracy and it also seems to catch those more intricate market movements, better than the alternatives. Overall, this framework can help with smarter investment decisions, portfolio oversight, and algorithmic trading, so it ends up being a workable approach for todays financial forecasting tasks.

Keywords: Financial Market Prediction, Hybrid LSTM-GRU, Deep Learning, Time-Series Forecasting, Technical Indicators, Artificial Intelligence, Machine Learning.

PROPOSED METHODOLOGY

1. Introduction

   The proposed methodology is basically gives a framework for trying to forecast financial market prices with a hybrid deep learning architecture, it mixes Long Short-Term Memory (LSTM) together with Gated Recurrent Unit (GRU) networks. Financial time-series data are known for volatility, nonlinear patterns, and this temporal dependency thing, so prediction is not that easy. To handle those aspects, the approach brings in historical market data preprocessing, some feature engineering, then sequence creation, next the hybrid model training and after that a performance evaluation. The workflow consists of six major phases:

   1. Historical Data Collection

   2. Data Preprocessing

   3. Feature Engineering

   4. Hybrid LSTMGRU Model Development

   5. Prediction

   6. Performance Evaluation

      Figure 6.1 Proposed Research Framework

      Historical Market Data

      Data Collection Module

      Data Cleaning & Validation

      Feature Engineering

      Sequence Generation (60 Days)

      Hybrid LSTM Layer

      GRU Layer

      Dense Layer

      Price Prediction

      Performance Evaluation

      Figure 6.1: Overall methodology for the proposed forecasting system.

2. Data Collection

   Historical market data are collected from publicly available financial data providers such as Yahoo Finance. Each record contains Open, High, Low, Close, and Volume values.

   Table 6.1 Dataset Description

   Attribute	Description	Data Type
   Date	Trading Day	Date
   Open	Opening Price	Float
   High	Highest Price	Float
   Low	Lowest Price	Float
   Close	Closing Price	Float
   Volume	Trading Volume	Integer

3. Data Preprocessing

   Raw financial data sets often have missing observations, duplicated entries, or weird scale

   differences, like sometimes numbers are all over the place. Before training, preprocessing helps, and yes it basically makes things cleaner.

   In the preprocessing flow youll usually see things like, Missing value removal, Duplicate removal, and Data validation. After that, they do MinMax normalization, then there is this step for sequence generation, where you turn the raw rows into ordered chunks for the model.

   Normalization Formula

   X=XXminXmaxXminX'=\frac{X-X_{min}}{X_{max}-X_{min}}X=XmaxXminXXmin

   where XXX is the original value, XminX_{min}Xmin and XmaxX_{max}Xmax are the minimum and maximum values of the feature.

   Figure 6.2 Data Preprocessing Flow

   Raw Dataset

   Missing Value Removal

   Duplicate Removal

   Normalization

   Sequence Generation

   Training Dataset

   Table 6.2 Preprocessing Activities

   Activity	Purpose
   Missing Value Handling	Improve data completeness
   Duplicate Removal	Eliminate redundant records
   Normalization	Standardize feature scales
   Sequence Generation	Prepare time-series inputs

4. Feature Engineering

   Feature engineering enhances the predictive capability of the model by combining raw market variables with derived indicators.

   Table 6.3 Input Features

   Feature	Description
   Open	Opening price
   High	Highest price
   Low	Lowest price
   Close	Closing price
   Volume	Trading volume
   SMA	Simple Moving Average
   EMA	Exponential Moving Average
   RSI	Relative Strength Index
   MACD	Moving Average Convergence Divergence

   Figure 6.3 Feature Engineering Process

   Historical Prices

   Technical Indicators

   Feature Selection

   Training Dataset

5. Hybrid LSTMGRU Architecture

   So, the forecasting model uses both LSTM and GRU bits. The LSTM part, helps it hold onto long- term ties in sequence like data, and it kind of keeps the context. Then the GRU layer comes in, and it lowers the computational load, plus it makes the training converge faster, maybe with less fuss.

   After that theres a Dense output layer that spits out the forecasted closing price. Table 6.4 Network Configuration

   Layer	Units	Activation
   LSTM	64	tanh
   GRU	32	tanh
   Dense	1	Linear

   Figure 6.4 Hybrid Network

   Input Sequence

   ++

   | LSTM Layer |

   ++

   ++

   | GRU Layer |

   ++

   ++

   | Dense Layer |

   ++

   Predicted Price

6. Model Training

   The normalized sequences are divided into training and testing sets. The model is trained using the Adam optimizer and Mean Squared Error (MSE) loss function. Hyper parameters such as learning rate, batch size, and number of epochs are selected through preliminary experiments.

   Table 6.5 Training Parameters

   Parameter	Value
   Epochs	20
   Batch Size	32
   Optimizer	Adam
   Loss Function	Mean Squared Error
   Sequence Length	60 Days

7. Performance Evaluation

   The trained model is evaluated using standard regression metrics.

   Table 6.6 Evaluation Metrics

   Metric Formula Purpose

   MAE Mean Absolute Error Average prediction error

   RMSE Root Mean Square Error Penalizes larger errors

   MAPE Mean Absolute Percentage Error Relative prediction error

   R² Score Coefficient of Determination Goodness of fit

8. Algorithm

   I collected historical financial data first. Missing and duplicate records were removed after that. Normalization was applied to the dataset so the values would sit on a similar scale. Time series sequences came from there.

   The hybrid LSTM GRU model got built next. Training used the training dataset. Predictions for future prices were then made with the testing dataset. Evaluation relied on MAE RMSE MAPE and R squared but some of the numbers felt stronger than others. I am not totally sure if that means the model handled everything evenly.

9. Chapter Summary

The proposed method for multi asset market prediction was laid out in this chapter. It mixes preprocessing with feature engineering and a hybrid LSTM GRU model to work with time patterns in the data.

That part feels like the core of it. The workflow tries to push accuracy higher and help with decisions but some details on how it all fits together still seem unclear in places.

REFERENCES

1. S. Hochreiter and J. Schmidhuber, "Long Short-Term Memory," Neural Computation, vol. 9, no. 8, pp. 17351780, 1997.

2. K. Cho et al., "Learning Phrase Representations using RNN EncoderDecoder for Statistical Machine Translation," EMNLP, 2014.

3. I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MIT Press, 2016.

4. D. P. Kingma and J. Ba, "Adam: A Method for Stochastic Optimization," ICLR, 2015.

5. B. Lim and S. Zohren, "Time-Series Forecasting with Deep Learning: A Survey," Philosophical Transactions of the Royal Society A, 2021.