0% found this document useful (0 votes)
26 views22 pages

Financial Market Prediction Using Deep Learning

Uploaded by

gy7q9nfgcr
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
26 views22 pages

Financial Market Prediction Using Deep Learning

Uploaded by

gy7q9nfgcr
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
You are on page 1/ 22

Financial Market Prediction Using Deep Learning

A Presentation Submitted
In Partial Fulfilment Of The Requirements
For The Degree Of

MASTER OF TECHNOLOGY
IN
DEPARTMENT OF
COMPUTER SCIENCE AND
ENGINEERING

Project Guide: Submitted By:


Mr. Surrender Yadav Nandini Kulshrestha
(23MCON0002)

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING


JECRC UNIVERSITY, JAIPUR
Problem Statement

The problem statement is to create prediction for


stock market using deep learning model. Simple
strategies like mean reversion, momentum
investing, systematic trend, etc. are popular in use.
Introduction

 Investment professionals utilize different ways to deal with oversee stock


portfolios. For this We will discuss two essential strategies that have been
around for quite a while — fundamental analysis and quantitative analysis.
 Fundamental Analysis :- Otherwise called "conventional" effective
financial investing, this approach utilizes a top to bottom analysis of an
organization's business, supervisory crew and market a potential open door
to decide the stock's engaging quality.
 Quantitative Analysis :- Otherwise called "structured" or "scientific"
investment management, this approach utilizes information driven analysis
to assess a wide universe of stocks.
Financial Data
Financial data is available in a variety of formats.
Tabular data (think spreadsheets) that can be
structured as rows and columns is the standard
format. Many websites, including finance.yahoo.com,
Quandl, Alpha Vantage, and many brokerages,
provide this type of information..
Financial data can be bought, manually scraped from
the web, or obtained from public APIs. Generally,
financial data comes in one of 2 primary types:
Structured Data: Closing prices, financials, market
performance, etc.
Unstructured Data: News articles, Social Media,
Sentiment Analysis, etc.
Market Data and Sources

 Markets trade related information is accessible in a variety of


organizations. Tables (spreadsheets, csv) that can be organized as lines and
sections is the standard configuration. Numerous sites, including
finance.yahoo.com, Quandl, Alpha Vantage, and various financiers, give
this sort of data.

 The information we are thinking about is from the date accessible of the
organization to 2018.
OHLC-V Data
Market data are commonly provided in a form which includes this information:
 Date: the date of day which is considered in data.
 Open Price: price at which trade open when market started.
 High Price: Highest price of a day.
 Low Price: Lowest price of a day
 Adjusting Closing Price: calculated closing price at which market close
 Volume: number of trades happened on that date, either amount of money
exchanged or number of trades.
The above data table represents the OHLC-V data from the Quandl data set

The company represented is ‘GOOGL’ - Alphabet Inc.


Sector wise stocks
Ticker IT sector Ticker Finance sector
AAPL Apple Inc BRK_B Berkshire Hathway Inc
GOOGL Alphabet Inc JPM JPMorgan Chase & Co
MSFT Microsoft corporation BAC Bank of America Corp
AMD Advanced Micro Devices, Inc C Citigroup Inc
INTC Intel Corporation WFC Wells Fargo & Co Ticker Communication services
NVDA Nvidia Corporation SPGI S&P Global Inc AMZN Amazon.com, Inc
CSCO Cisco Systems Inc MS Morgan Stanley TSLA Tesla Inc
ADBE Adobe Inc. HD Home Depot Inc
MCD McDonald's Corp
Ticker Consumer discretionary Ticker health
SBUX Starbucks Corporation
GOOG Alphabet Inc UNH UnitedHealth Group Inc
NKE Nike Inc
FB Meta Platforms Inc JNJ Johnson & Johnson
LOW Lowe`s Companies Inc
VZ Verizon Communications Inc ABBV AbbVie Inc
DIS Walt Disney Co MRK Merck & Co., Inc
T AT&T Inc TMO Thermo Fisher Scientific Inc
NFLX Netflix Inc ABT Abbott Laboratories
CMCSA Comcast Corporation PFE Pfizer Inc
TMUS T-Mobile Us Inc LLY Eli Lilly And Co
price relative to maximum price
Price relative to maximum price
Data handling
Data Pull: Market data for relevant stocks was pulled
from Quandl python API
Data Selection: Out of all stocks in NASDAQ, the ones which were
traded in last 30 days, and has more than 10 years of running data are
taken into account
Data Cleaning: Although Quandl provides us with adjusted price,
Null value analysis and handling for some companies needs to be
done, for days when exchange was open but particular stock was not
traded due to any reason (including but not limited to legal issues,
sanctions, etc.,) Null spaces are filled using Forward fill method
provided in Pandas library
Feature Normalization: OHLC-V raw data for each stock can vary
in scale, making it difficult for a model to learn. Feature normalisation
helps brings every value in same scale, dividing by last or max value
of time series
Pre-processing and normalization
technique
• Our pre-processing technique
ensures last observed prices is of
unit value.

• This ensure for any sequential


model only relative change in price
needs to be predicted in respect to
the actual historical price.
methodology

Model Training: All models which we have experimented with are variants
of Neural networks, which are training to minimise differentiable loss (Mean
Squared Error). Adam was used as an optimiser which is a de-facto standard
to achieve faster convergence and better accuracies. Optimizer parameter are
as below: learning rate = 0.0001 batch size = 128

Regression Evaluation: Evaluation is done using MAPE loss which is


relevant to Trader, also to account for prediction in price movement to
establish model confidence. The mean absolute percentage error (MAPE), also
known as mean absolute percentage deviation (MAPD). Also price movement
accuracy was assigned when the price movement of the stock was in the same
direction as suggested by the model.
Our contribution
Data normalization
 It helps in bringing data on to the same scale
 It can be applied to the entire data at one or apply at batch level.
 Since our batch cover is 90 days of data , and the entire time series spanmore than ten years. Hence
batch level technuiqe are more suitable for us.

 We developed last point normalization which brings last data to unit value, bringing all next day price
prediction close to one.
result
It sector
result
Health sector
result
Communication
Services
Sector
comparison
Predicting a reliable stock for mid- and long-term investment, YMCA vs our
approch

1. Investing, thus having limited chance 1. A huge volume of suitable trade over
of beating market as investment will a short period of time, thus
be of mid and large duration. False- generating profit for long term is
positive must be low. maximized over large number of
trades (LLN)
2. Humans decide thus higher chances
of biases and slower execution speed. 2. Automatic provides no emotional
biases and much faster execution
3. Not possible as expert opinion keeps
speed
changing with time. Could be
intuition-based. 3. Yes, by having required resources
one could easily do it.
4. Is the main factor for most of the
decision making 4. Only limited to maths and related
data science.
A Novel Graph Convolutional Feature Based Convolutional Neural Network for
Stock Trend Prediction vs Our Proposed Model

A Novel Graph Convolutional Feature Based Convolutional Neural Our Proposed Model
Network for Stock Trend Prediction

Correlation among stock prices were used to pick top related companies Companies were taken dependent on volume and market popularity and
independent of movement correlation
Market data was converted in form of images Raw market data was used in form of OHLCV

Model was highly complex which can possibly take all form of features Model internally learns only required feature representation, preventing
which can be created using raw data chances of overfitting

Showed study on 6 stock prices only Considered set of 38 stock prices

Performed study on different networks, but looks like deliberately skipped Obtained good performance from the model which study of other authors
CNN model applied on RAW data skipped

Results were shown for different time period in month (wait unless the model Same time frame for all models and all stock prices were used
performance beats others)
Advantages of our approach

 Trades are done, best case scenario, potential costs.


 Trade demand circumstance is moment and exact (there is a high chance of
execution at the ideal levels consequently staying away from low liquidity
or huge spread trades).
 Diminished danger of manual missteps or human carelessness while
exchanging.
 Algo-trading can be back-tried on verifiable and live information to check
regardless of whether it is appropriate for exchanging
References

 https://jfin-swufe.springeropen.com/articles/10.1186/s40854-019-0131-7
 Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow, 2nd Edition by
Aurélien Géron
 Fausett L (1994) Fundamentals of neural networks. Prentice Hall, New York, NY, USA
 McKinney, Wes. 2017. Python for Data Analysis. 2nd ed. Sebastopol: O’Reilly.
 VanderPlas, Jake. 2016. Python Data Science Handbook. Sebastopol: O’Reilly.
 Listed Volatility and Variance Derivatives: A Python-Based Guide. Wiley Finance.
 Python for Finance: Mastering Data-Driven Finance. 2nd ed. Sebastopol: O’Reilly
Thank You

You might also like