Skip to content

The learnings of the members of team CryptoFutureX during SkillUp 2.0

Notifications You must be signed in to change notification settings

CryptoFutureX/learning

Repository files navigation

CryptoFutureX logo CryptoFutureX

Team members

  • Pankaj Khushalani
  • Tushar Bauskar
  • Utsav Khatu
  • Nikita Deokar

Mentors

  • Sanket Barhate
  • Dhwani Panjwani
  • Suyash Gatkal

Description

A web application to view futures of various cryptocurrencies.
The predictions are made by using the trends of prices of cryptocurrencies and correlating them with sentiment analysis. Multivariate time series analysis is used to predict the trends and the percentage change in predicted trends is obtained by sentiment analysis of Twitter feed.

  • On the ReactJS frontend, user can view candlebar plot of prices of selected cryptocurreny over the past 7 days.

  • The futures of the 8th day are predicted and shown in the same plot. The prices can be viewed by toggling the selection in USD, INR, EUR, and JYP.

  • Percentage change in predicted price by sentiment is indicated by a speedometer below the graph.

GitHub organization link: CryptoFutureX

Google Drive folder: Link

Installation

  • For the frontend,

    • Clone the repository

      git clone https://github.com/CryptoFutureX/frontend.git
      cd frontend/client
      npm install
    • Get your API key from CoinLayer. Create key.js and add your API key

      API_KEY = <your-API-key>
    • Then start the server on port 3000 by

      npm start
  • For the backend,

    • Clone the repository
      git clone https://github.com/CryptoFutureX/backend.git
      cd backend
    • Create a virtual environment using venv (for Windows)
      python3 -m venv <your-virtual-env-name>
      .\env\Scripts\activate
    • Run the backend server on port 4000 by
      python3 manage.py runserver

    Since the packages to install require several C++ dependencies and PyTorch along with TensorFlow are heavy libraries, an alternative is using Google Colab.

    • Open a Google Colaboratory notebook and mount your Google Drive.
      from google.colab import drive
      drive.mount('/content/drive/MyDrive')
    • Now, clone the repository in your Google Drive and install any dependencies which might not already be installed on Google Colab.
      !git clone https://github.com/CryptoFutureX/backend.git
      %cd backend
      !pip install -r requirements.txt
    • Now expose a port to Google Colab in order to run the Django app
      from google.colab.output import eval_js
      print(eval_js("google.colab.kernel.proxyPort(4000)"))
    • In settings.py
      ALLOWED_HOSTS = ['colab.research.google.com']
    • Finally, run the server on port 4000
      !python manage.py runserver 4000

Technology stack

  1. Django
    A Python framework for web application development. For the backend, Django with Django Rest Framework was used to supply data to the ReactJS frontend.

  2. ReactJS
    A JavaScript library for building user interfaces. For data visualization, Apex charts were used along with React components.

  3. TensorFlow
    A Python library for machine learning. Keras API was used along with TensorFlow to train the CNN and LSTM models.

  4. APIs Used

    • Twint - An advanced Twitter scraping tool written in Python that allows for scraping Tweets from Twitter profiles without using Twitter's API.

    • CryptoCompare - It is a Python library for fetching prices both latest as well as historical, coin listing across various exchanges and also trading using the API.

  5. Dataset

    • Kaggle dataset containing 1.4 million BitCoin tweets collected in 2 weeks of July 2018 with their sentiment. This was used for the baseline model. However, the tweets had emojis with their raw UTF-8 representations, making it tough to clean the data.

    • Sentiment140 dataset containing tweets of 140 categories with their respective sentiment. This is a generalized data for Twitter sentiment analysis created by Stanford and was used to train our final model.

Applications

Barring a stock exchange of cryptocurrencies, there is considerable risk in investment in any cryptocurrency. With CryptoFutureX, an investor can visualize the predicted change in prices and make the right investment.

Skills Learnt

We explored several technologies pertaining to the field of Natural Language Processing (NLP). Team members explored alternatives to few technologies while some were learnt by all.

Common technologies

  1. NLTK (Natural Language Toolkit)
    It is a prevalent NLP library in Python which has rich sources of human language data. We learnt concepts such as tokenization, lemmatization, stemming, POS tagging, BOW and n-grams using NLTK.
    However, with unavailibilty of vector embeddings and neural network models, other alternatives were explored. References:

  2. PyTorch
    It is is an optimized tensor library for deep learning using GPUs and CPUs.
    We chose this library as it has rich implementations in NLP such as transformers library by Hugging Face. It provides thousands of pre-trained models for NLP such as BERT, RoBERT, DistilBERT. BERT was extensively used in training our model.
    References:

  3. TensorFlow
    It is a Python library used for machine learning. We used TensorFlow along with Keras API to do multivariate time series modelling using LSTM (an artificial recurrent neural network) for price prediction.
    References:

Technologies learnt by team members

  • Pankaj Khushalani
    spaCy - Though NLTK is great for a beginner, it didn't provide us with the necessary tools to support word vector embeddings and to build our own model. Hence, spaCy, an advanced NLP library was used.
    Being compatible with classes from scikit-learn, building models with spaCy becomes easy. Following were the ML algorithms used with TF-IDF vectorized data:

    ML Algorithm Train acc. Test acc.
    LinearSVC 0.830495 0.762606
    SGDClassifier 0.758569 0.755504
    DecisionTree 0.6212434 0.6210907
    MultinomialNB 0.7971801 0.7562758

    Though spaCy along with support vector machines gave good accuracy, the CNN model gave more reliable and reproducible results.

  • Tushar Bauskar
    CNN model for sentiment analysis
    As we were not able to use BERT model for sentiment analysis due to large data and less computational power,a CNN model was created. CNN can be used for sentiment analysis by creating vector embeddings of the text. The vectors were created using pretrained GloVe Twitter Embedding 27B 100D vector embeddings. It contains 27 billion words and their embeddings in 100 dimension vectors.

    CNN Model

    class CnnModel(nn.Module):
        def __init__(self):
            super().__init__()
            self.vocab_size = 6002
            self.embed_size = 100
            self.num_filters = 100
            self.filter_sizes = [3, 4, 5]
            self.output_classes = 2
            self.dropout = 0.8
    
            # Embedding layer
            self.embedding = nn.Embedding(self.vocab_size, self.embed_size)
    
            # Convolutional layer
            self.convs = nn.ModuleList([
                                    nn.Conv2d(
                                        in_channels=1, 
                                        out_channels=self.num_filters,
                                        kernel_size=(fs, self.embed_size)) 
                                    for fs in self.filter_sizes
            ])
    
            # Fully connected layer
            self.fc = nn.Linear(len(self.filter_sizes) * self.num_filters, self.output_classes)
    
            # Dropout layer
            self.dropout = nn.Dropout(self.dropout)
    
        def forward(self, text, text_lengths):
            embedded = self.embedding(text)
            embedded = embedded.unsqueeze(1)
    
            conved = [F.relu(conv(embedded)).squeeze(3) for conv in self.convs]
            pooled = [F.max_pool1d(conv, conv.shape[2]).squeeze(2) for conv in conved]
    
            cat = self.dropout(torch.cat(pooled, dim=1))
            return self.fc(cat)


We used the Sentiment140 dataset for training the model.
Accuracy achieved:

Training accuracy Validation accuracy
81.52% 81.16%

  • Utsav Khatu

    • D3JS: We initially decided to use D3Js for the frontend to create and plot prcice but soon realized that the D3Js is a low level library making it easier to customize according to the users but even the small changes takes a lot of effort to create and soon switched to ChartJs

    • ChartJS: Next we stumbled on ChartJs since we initially decided to plot a line plot so it was rather easy since ChartJs is Simple, clean and engaging HTML5 based JavaScript charts.

    • React ChartJS: Since ChartJs was mainly used for vanillaJs we needed a library which could configure and can be updated using React. This is where react-chart library was used.

    • React ApexCharts: Later we decided to give more investor oriented look to our webapp and decided to implement Candle Plot & Line Plot in our webapp.This is where ApexCharts Library came in handy.

  • Nikita Deokar

    • BERT stands for Bidirectional Encoder Representations from Transformers. It is pre-trained on a large corpus of unlabelled text including the entire Wikipedia and Book Corpus. Thus, BERT has proven to be a benchmark model for many NLP models.
      However, due to a huge dataset and low computational power a BERT model could not be trained. To address these problems, we explored DistilBERT and ALBERT variations of BERT.

    • DistilBERT has 40% lesser parameters than bert-base-uncased, runs 60% faster while preserving over 95% of BERT’s performance.

    • ALBERT is a lite representation of BERT that reduces energy consumption while increasing training speed compared to BERT model.


However, none of the above deep learning transformer models could be trained with the freely available resources such as Google Colab or the free tier usage on GCP or AWS. Thus, results could not be produced from any of the above models.

Future Scope

  • Automating the data collection and prediction pipeline on the backend.
  • Improving the accuracy of the current model and exploring other NLP techniques to get better predictions.
  • Adding live indicator of cryptocurrency prices along with the candlebar graph.
  • Deployment of the data collection script to AWS Lambda and price prediction model to AWS SageMaker.

Screenshots

Frontend chart

Frontend price change by sentiment

About

The learnings of the members of team CryptoFutureX during SkillUp 2.0

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published