Return prediction by

Machine learning for

Indian Stock Market
Rahul Sharma
Date of Submission: 14 Oct 2024

Fig 1. Types of Models used in Stock Forecasting

1. Introduction
Predicting the future of the Indian stock market using machine learning has been a topic of
increasing interest. "Return prediction by machine learning" involves using advanced
algorithms to forecast stock price movements based on historical data. This helps investors
make more informed decisions regarding buying and selling stocks.

The Indian stock market is known for its volatility, and having a reliable tool to forecast
these fluctuations can be highly beneficial. Machine learning models, such as those based
on historical stock prices and other influential factors, attempt to predict the market's
direction. Understanding and leveraging machine learning can provide investors with a
competitive edge. This paper explores the application of machine learning to stock market
predictions, focusing on specific techniques like ARIMA, LSTM, and CNN.

Fama (2024) introduced the Efficient Market Hypothesis, asserting that current stock prices
fully reflect all available information. Additionally, the Random Walk Hypothesis posits that
stock prices change independently of their history, making accurate predictions theoretically
impossible. However, contrary to these ideas, other researchers argue that stock prices can
be predicted to some extent, utilizing various disciplines like economics, statistics, and
computer science (Lo & McKinlay, 1999).

Fig. 2 Structure of an LSTM unit. Source: Ding et al. (2015).

Given the intrinsic complexity of the stock market, there has always been discussion over
how predictable stock returns are, and this topic has generated a lot of research. According
to Fama (2024), the efficient market hypothesis states that an asset's current price always
accurately represents all previous information that was instantly available to it. Furthermore,
according to the random walk hypothesis, a stock's price fluctuates independently of its past
performance; in other words, the price of a stock tomorrow will only be determined by
information available tomorrow, not by the price of a stock today (India, 2024)). These two
theories prove that it is impossible to predict stock prices with any degree of accuracy.
Some authors, on the other hand, contend that stock values can be forecast, at least
somewhat. and a range of techniques

2. Methodology
This section explains the methodology used for stock price prediction using Python libraries
and machine learning techniques.

2.1 Python Libraries

Several Python libraries are employed for data analysis and model
implementation:

 Pandas: For loading and managing data in a tabular format.

 Numpy: For efficient computation and handling large datasets.
 Matplotlib/Seaborn: For data visualization.
 Sklearn: A suite of libraries for data preprocessing, model building, and evaluation.
 XGBoost: A powerful machine learning algorithm that provides high prediction
accuracy.

2.2 ARIMA Models

ARIMA (AutoRegressive Integrated Moving Average) models are commonly used for time
series analysis and forecasting. This study aims to find the optimal p, d, and q parameters
for ARIMA, which are critical for improving forecasting accuracy.

Context and Associated Works

Because ARIMA models can manage trends and seasonality, they are frequently employed in
time series analysis. Various approaches to parameter selection have been studied in the
past, such as grid search and optimization algorithms. But in this work, we systematically
examine different combinations of p and q values using a nested loop approach.

Techniques
The goal variable for forecasting is a time series dataset called y. We use a nested loop
technique to get the ideal p and q parameters. Predefined lists of potential values for p and
q are called p_params and q_params, respectively. The ARIMA function is used to fit an
ARIMA model with the order (p, 0, q) to the data for each combination of p and q.

CNN and LSTM

Deciphering the underlying temporal patterns within data sequences is the basis of
sequence prediction, which is used in speech recognition, financial forecasting, and natural
language processing, among other applications. This work is extremely important because it
gives decision-makers the capacity to foresee trends in the future, modify their approaches,
and make well-informed decisions. Advances in sequence modeling have been accelerated
by the use of Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs)
in deep learning, which have shown to be adept at processing sequential and grid-like data,
respectively.

Initialization of the Model

The class is used to initialize a sequential model, enabling the building of a linear stack of
layers. Using sequential data, this model will be used to identify trends and forecast
outcomes.

Conv1D Layer with Time-Distributed

An additional TimeDistributed layer is added to the model in order to capture temporal
interdependence within the sequential data. In order to extract local features, this layer has
a Conv1D layer with 64 filters and a kernel size of 1. The model is made non-linear by
applying the rectified linear unit (ReLU) activation function. of the {input_shape} option set
to {(None, 1, time_stemp)}, the model is capable of handling input sequences of different
lengths.

Time Distributed MaxPooling1D Layer

To down sample the feature maps obtained from the previous layer, a Time Distributed layer
with a MaxPooling1D layer is incorporated into the model. The MaxPooling1D layer has a
pool size of 2 and employs 'same' padding to maintain the spatial dimensions of the feature
maps.

Time Distributed Flatten Layer

A Time Distributed layer is introduced to perform the flattening operation on the output of
the previous layer. This operation transforms the 3D feature maps into a 2D representation,
which facilitates further processing.

LSTM Layer
To capture long-term dependencies and sequential patterns, an LSTM layer with 50 units is
added to the model. The rectified linear unit (ReLU) activation function is employed within
the LSTM layer to introduce non-linearities. This layer processes the sequential data from
the previous Time Distributed layers and extracts relevant features.

Dense Layer
Following the LSTM layer, a single-node dense layer is included in the model to produce the
desired output. This layer uses a linear activation function, which ensures that the output
can span a wide range of values.

Compilation
To configure the model for training, it is compiled using the Adam optimizer. The mean
squared error (MSE) loss function is selected, which measures the discrepancy between the
predicted and true values. The Adam optimizer is known for its efficiency in training deep
neural networks and is widely used in various domains.

3. Results
In this section, we present the results of the machine learning models, including ARIMA,
LSTM, and CNN. Each model's accuracy, represented by the Mean Absolute Error (MAE), is
calculated and compared. ARIMA's effectiveness in capturing time series patterns and
LSTM's ability to retain long-term dependencies are highlighted.

Training
The model is trained using the `fit` function, which applies the backpropagation algorithm to
update the model's weights iteratively. The training data, denoted as `trainX` and `trainY`,
represent the input sequences and their corresponding target values, respectively. The
training process consists of 50 epochs, indicating that the entire training dataset is
traversed 50 times. A batch size of 1 is used, meaning that a single sample is processed in
each iteration. This choice allows for fine-grained updates of the model's parameters and
can be beneficial when dealing with sequential data.

By following this model architecture and training procedure, the model learns to capture
temporal patterns and make accurate predictions on sequential data. The utilization of
convolutional and recurrent layers enables the model to extract local features and capture
long-term dependencies, respectively, resulting in an effective framework for sequential
data analysis.

Gated Recurrent Unit (GRU)

Time series prediction, a pivotal task in predictive analytics, holds immense significance
across diverse domains such as finance, healthcare, climate prediction, and industrial
maintenance. Accurate forecasting of sequential data enables informed decision-making and
proactive measures, leading to improved outcomes and resource allocation. In recent years,
Recurrent Neural Networks (RNNs) have emerged as powerful tools for modelling sequential
data due to their inherent ability to capture temporal dependencies. Among the RNN
variants, the Gated Recurrent Unit (GRU) stands out as a compelling architecture, known for
its simplified gating mechanism and memory management. GRUs exhibit remarkable
performance in learning sequential patterns and have gained traction in various
applications.
Stock Price Prediction using Machine Learning
In this project, we aim to predict whether buying Tesla stock will be profitable using machine
learning. We'll go through the process of data importation, exploratory data analysis, feature
engineering, model development, and evaluation.

1. Import the Libraries.

2. Load the Training Dataset.

The Google training data has information from 3 Jan 2012 to 30 Dec 2016. There are five
columns. The Open column tells the price at which a stock started trading when the market
opened on a particular day. The Close column refers to the price of an individual stock when
the stock exchange closed the market for the day. The High column depicts the highest
price at which a stock traded during a period. The Low column tells the lowest price of the
period. Volume is the total amount of trading activity during a period of time

.
3. Use the Open Stock Price Column to Train Your Mode4.
Normalizing the Dataset.

4. Normalizing the Dataset.

5. Creating X_train and y_train Data Structures.

6. Reshape the Data.

7. Building the Model by Importing the Crucial Libraries
and Adding Different Layers to LSTM.

8. Fitting the Model.

9. Extracting the Actual Stock Prices of Jan-2024.

10. Preparing the Input for the Model.

11. Predicting the Values for Jan 2017 Stock Prices.

12. Plotting the Actual and Predicted Prices for Google

Stocks.
(Code is written by myself and is not taken by any references)

As you can see above, the model can predict the trend of the actual stock prices very
closely. The accuracy of the model can be enhanced by training with more data and
increasing the LSTM layers.
4. Discussion
The results show that the choice of parameters, particularly for ARIMA models, has a
significant impact on forecasting accuracy. Larger values of p and q can cause overfitting,
while smaller values may lead to underfitting. Moreover, the LSTM model's performance
improves with additional layers and longer training times, capturing the temporal
dependencies more effectively.

 Normalization: Normalization helps in stabilizing the training process by scaling the

feature values.
 Data Splitting: The dataset is split into training and validation sets to ensure the
model is evaluated on unseen data.
 LSTM Layers: LSTM is chosen due to its ability to capture long-term dependencies in
time-series data.
 Plotting: Visualizing predictions against actual values helps in understanding model
performance and identifying any issues.

Bidirectional LSTM (BiLSTM)

In recent years, the Bidirectional LSTM architecture has gained prominence as a compelling
extension to the standard LSTM. This architecture augments the LSTM cell by allowing
information flow not only in the forward temporal direction but also in the reverse direction.
By capturing dependencies in both directions, the Bidirectional LSTM can uncover contextual
cues that may remain hidden in unidirectional models. This innovation has found
applications in diverse fields including natural language processing, speech recognition, and,
notably, time series analysis.

Model Initialization
A sequential model is initialized using the Keras framework's `Sequential` class. This model
allows for the construction of a linear stack of layers, facilitating the definition of deep
learning architecture.

Bidirectional LSTM Layer

To capture the temporal dependencies and patterns within the input data, a bidirectional
Long Short-Term Memory (LSTM) layer is employed. The LSTM layer consists of 50 units and
uses the rectified linear unit (ReLU) activation function. The bidirectional nature of this layer
allows it to process the input data in both the forward and backward directions, enabling a
comprehensive

Dense Layer
Following the bidirectional LSTM layer, a dense layer with a single node is added to the
model. This layer is responsible for producing the output of the model. The use of a single-
node dense layer suggests that the regression task aims to predict a single continuous value
as the output.

Compilation
To prepare the model for training, it is compiled using the Adam optimizer. The Adam
optimizer is a popular choice for deep learning tasks due to its adaptive learning rate and
efficient optimization capabilities. The mean squared error (MSE) loss function is employed,
which measures the discrepancy between the predicted and true values. By minimizing this
loss function, the model aims to learn to accurately predict the target values.

Training
The model is trained using the `fit` function, which applies the backpropagation algorithm to
update the model's weights iteratively. The training dataset, denoted as `trainX` and
`trainY`, represents the input sequences and their corresponding target values, respectively.
The training process consists of 50 epochs, indicating that the entire training dataset is
traversed 50 times during training. A batch size of 1 is chosen, meaning that each sample is
processed individually, allowing for fine-grained updates of the model's parameters.

By utilizing a bidirectional LSTM layer and a single-node dense layer, the model learns to
capture and understand the temporal dependencies within the input data, ultimately
predicting a continuous value for the given regression task. The training process, consisting
of multiple epochs and fine-grained updates with a batch size of 1, enables the model to
optimize its parameters and improve its predictive performance.

ARIMA Model
This research also explores the process of building and optimizing ARIMA (Auto Regressive
Integrated Moving Average) model parameters for time series forecasting. The study
involves experimenting with different combinations of p and q parameters to identify the
most accurate model. The Mean Absolute Error (MAE) is employed as a performance metric
to evaluate the accuracy of each model. The results highlight the significance of selecting
appropriate parameters for achieving better forecasting accuracy in time series analysis.

5. Conclusion
This paper demonstrates that while predicting stock prices is inherently challenging,
machine learning models such as ARIMA and LSTM show promise in achieving reasonable
accuracy. By further tuning these models and using more extensive datasets, prediction
accuracy can be enhanced.

REFERENCES:
 Machine learning approaches in stock market prediction: A systematic literature
review - ScienceDirect
 Electronics | Free Full-Text | Stock Market Prediction Using Machine Learning
Techniques: A Decade Survey on Methodologies, Recent Developments, and Future
Directions (mdpi.com)
 Pooling and Winsorizing Machine Learning Forecasts to Predict Stock Returns with
High-Dimensional Data - ScienceDirect
 (PDF) Stock Prediction Using Machine Learning (researchgate.net)
 Fama, E. (2024). Efficient Market Hypothesis. [Publisher details].
 Lo, A.W., & McKinlay, A.C. (1999). A Non-Random Walk Down Wall Street. Princeton
University Press.