Deep Learning Course 046211 Project: Predictinting NASDAQ-100 using LSTM&GRU and Federal Reserve data
Python implementation for NASDAQ-100 price prediction using FED's assets.
Project has been done by Roy Mahlab & Tamir Haver.
In this project we coded in python using Jupyter Notebook.
When runnig the code on Jupyter one should run it cell by cell in the same order as it is brought in our repository.
The Federal Reserve System (also known as the Federal Reserve or simply the Fed) is the central banking system of the United States of America.
Since the corona-virus pandemic broke, the United States Federal Reserve assets nearly doubled, which raised claims that the Fed is inflating the value of assets in the stock market.
In our project we will examine this claim. We will do so by implementing 2 different methods for RNN models, and compare their results with each other and examine the influence of the FED's assets on the results.
We can divide the workflow into few main steps:
Loading the data
Spliting the data
Building the LSTM and GRU models
Using OPTUNA for tuning of hyper parameters
Compering the results and the naive prediction (average weekly gain)
Presenting the results
The hyper parameters we used OPTUNA for tuning are:
Batch Size
Hidden and Output Dim of GRU and LSTM
N layers of GRU and LSTM
Learning Rate
Look Back – How many past weeks do we need in order to predict next week’s value.
We compared our models to a naïve predictor which predicts next week’s closing value based on the average weekly gain – g of the Nasdaq:
| File Name | Purpose |
|---|---|
| Prediction_FED.ipynb | Main code file including all the projects results |
| optuna_search.py | Hyper parameters tuning |
| utils.py | Loading the data, declaring project's models,models training, declaring obective for optuna and printing functions |
| studies | Optunas' studies results |
| input | The data we used for our project which includes Nasdaq-100 index values and the FED assets |
in order to run using existing studies one can run Prediction_FED.ipynb. You can also run optuna_search.py in order to create a new study, and adjust parameters within it.
We shall denote, t - the target closing price, p - our prediction for the closing price. The loss over a batch will be the MAPE – mean absolute percentage error:
And so:
Future prediction (when the last data dates back to June 23rd):
Fed total assets: https://fred.stlouisfed.org/series/WALCL
Based on the work "Predicting Stock Price using LSTM model, PyTorch" by Taron Zakaryan, Paris, Île-de-France, France.
Project site can be found here: https://www.kaggle.com/taronzakaryan/predicting-stock-price-using-lstm-model-pytorch