Forecasting FX rates with LSTM networks using the RNN SageMaker algorithm¶
Forecasting exchange rates is a critical task for traders, investors, and financial institutions. Even though different machine learning models have been studied for this purpose, Long Short-Term Memory (LSTM) [1] networks have become the most widely adopted [2].
LSTMs belong to the class of recurrent neural networks (RNNs), which are designed to process and predict sequences of data. Differently from vanilla RNNs, which often fail to capture long-term dependencies due to issues like vanishing gradients, LSTMs implement a number of gating mechanisms which allow the network to retain memory of relevant features over longer time intervals.
In this post, we will use our Amazon SageMaker implementation of RNNs for probabilistic time series forecasting, the RNN SageMaker algorithm. We will implement a stacked LSTM model with two layers with LeCun's hyperbolic tangent activation [3]. We will use the model for forecasting the EUR/USD exchange rate one day ahead using as input a set of technical indicators, similar to [4].
We will use the daily EUR/USD exchange rate time series from the 1st of August 2022 to the 31st of July 2024, which we will download from Yahoo! Finance. We will train the model on the data up to the 18th of June 2024, and use the trained model to predict the subsequent data up to the 31st of July 2024. We will find that the model achieves a mean absolute error of 0.0012 and a mean directional accuracy of 83.33% over the considered time period.
Code¶
Warning
To be able to run the code below, you need to have an active subscription to the RNN SageMaker algorithm. You can subscribe to a free trial from the AWS Marketplace in order to get your Amazon Resource Name (ARN). In this post we use version 1.0 of the RNN SageMaker algorithm, which runs in the PyTorch 2.1.0 Python 3.10 deep learning container.
Environment Set-Up¶
We start by importing all the dependencies and setting up the SageMaker environment.
Note
We use the yfinance library for downloading the EUR/USD daily time series and
the pyti library for calculating the technical indicators.
import io
import sagemaker
import warnings
import yfinance as yf
import pandas as pd
import numpy as np
import statsmodels.api as sm
import matplotlib.pyplot as plt
from pyti.momentum import momentum
from pyti.rate_of_change import rate_of_change
from pyti.moving_average_convergence_divergence import moving_average_convergence_divergence
from pyti.simple_moving_average import simple_moving_average
from pyti.relative_strength_index import relative_strength_index
from pyti.bollinger_bands import middle_bollinger_band, upper_bollinger_band, lower_bollinger_band
from pyti.commodity_channel_index import commodity_channel_index
from sklearn.metrics import root_mean_squared_error, mean_absolute_error, accuracy_score, f1_score
warnings.filterwarnings(action='ignore')
# SageMaker algorithm ARN, replace the placeholder below with your AWS Marketplace ARN
algo_arn = "arn:aws:sagemaker:<...>"
# SageMaker session
sagemaker_session = sagemaker.Session()
# SageMaker role
role = sagemaker.get_execution_role()
# S3 bucket
bucket = sagemaker_session.default_bucket()
# EC2 instance
instance_type = "ml.m5.4xlarge"
After that we define the neural network’s context length and prediction length. The context length is the number of past time steps used as input, while the prediction length is the number of future time steps to be predicted. We set the context length equal to 5 and the prediction length equal to 1, that is we use the values of the EUR/USD exchange rate and of the technical indicators on the previous week to predict the value of the EUR/USD exchange rate on the next day.
# number of time steps used as input
context_length = 5
# number of time steps to output
prediction_length = 1
Data¶
Next, we download the EUR/USD exchange rate time series from the 1st of August 2022 to the 31st of July 2024. The dataset contains 522 daily observations.
dataset = yf.download(tickers="EURUSD=X", start="2022-08-01", end="2024-08-01")
We then calculate the following technical indicators, as in [4]:
Moving average (MA) with a period of 10.
Moving average convergence/divergence (MACD) with periods of 12 and 26.
Rate of change (ROC) with a period of 2.
Momentum with a period of 4.
Relative strength index (RSI) with a period of 10.
Bollinger bands (BB) with period of 20.
Commodity channel index (CCI) with a period of 20.
# MA with a period of 10
dataset["MA"] = simple_moving_average(
data=dataset["Close"],
period=10
)
# MACD with short- and long-term periods of 12 and 26
dataset["MACD"] = moving_average_convergence_divergence(
data=dataset["Close"],
short_period=12,
long_period=26
)
# ROC with a period of 2
dataset["ROC"] = rate_of_change(
data=dataset["Close"],
period=2
)
# Momentum with a period of 4
dataset["Momentum"] = momentum(
data=dataset["Close"],
period=4
)
# RSI with a period of 10
dataset["RSI"] = relative_strength_index(
data=dataset["Close"],
period=10
)
# BB with period of 20
dataset["MiddleBB"] = middle_bollinger_band(
data=dataset["Close"],
period=20
)
dataset["LowerBB"] = upper_bollinger_band(
data=dataset["Close"],
period=20
)
dataset["UpperBB"] = lower_bollinger_band(
data=dataset["Close"],
period=20
)
# CCI with a period of 20
dataset["CCI"] = commodity_channel_index(
close_data=dataset["Close"],
low_data=dataset["Low"],
high_data=dataset["High"],
period=20
)
After dropping the missing values resulting from the calculation of the technical indicators, the number of daily observations is reduced to 497.
# drop the missing values
dataset.dropna(inplace=True)
# drop the unnecessary columns
dataset.drop(labels=["Adj Close", "Volume"], axis=1, inplace=True)
# move the target to the first column
dataset = dataset[["Close"] + dataset.columns.drop("Close").tolist()]
dataset.shape
(497, 13)
dataset.head()
dataset.tail()
EUR/USD daily exchange rate with technical indicators from 2022-09-05 to 2024-07-31.
We now proceed to renaming the columns in the format expected by the RNN SageMaker algorithm,
where the output names should start with "y" and the input names should start with "x".
# rename the columns
dataset.columns = ["y"] + [f"x{i}" for i in range(dataset.shape[1] - 1)]
dataset.head()
dataset.tail()
Note
Note that the algorithm’s code always includes the past values of the outputs among the inputs and, therefore, there is no need to add the lagged values of the outputs when preparing the data for the model.
We then split the data into a training set and a test set. We use the last 30 days for testing, and the previous 467 days for training. We save both the training data and the test data to CSV files in S3 such that they can be used by the algorithm.
# define the size of the test set
test_size = 30
# extract the training data
training_dataset = dataset.iloc[:- test_size]
# extract the test data
test_dataset = dataset.iloc[- test_size - context_length:]
# upload the training data to S3
training_data = sagemaker_session.upload_string_as_file_body(
body=training_dataset.to_csv(index=False),
bucket=bucket,
key="training_data.csv"
)
# upload the test data to S3
test_data = sagemaker_session.upload_string_as_file_body(
body=test_dataset.to_csv(index=False),
bucket=bucket,
key="test_data.csv"
)
Note
Note that the data is scaled internally by the algorithm, there is no need to scale the data beforehand.
Training¶
We can now train the model using the training data in S3. We use two LSTM layers with respectively 100 and 50 hidden units and apply a LeCun's hyperbolic tangent activation [3] after each layer. We train the model for 200 epochs with a batch size of 16 and a learning rate of 0.001, where the learning rate is decayed exponentially at a rate of 0.99.
# create the estimator
estimator = sagemaker.algorithm.AlgorithmEstimator(
algorithm_arn=algo_arn,
role=role,
instance_count=1,
instance_type=instance_type,
input_mode="File",
sagemaker_session=sagemaker_session,
hyperparameters={
"context-length": context_length,
"prediction-length": prediction_length,
"sequence-stride": 1,
"cell-type": "lstm",
"hidden-size-1": 100,
"hidden-size-2": 50,
"hidden-size-3": 0,
"activation": "lecun",
"dropout": 0,
"batch-size": 16,
"lr": 0.001,
"lr-decay": 0.99,
"epochs": 200,
}
)
# run the training job
estimator.fit({"training": training_data})
Inference¶
After the training job has been completed, we run a batch transform job on the test data in S3.
The results of the batch transform job are saved to a CSV file in S3 with the same name as the
input CSV file but with the ".out" file extension.
# create the transformer
transformer = estimator.transformer(
instance_count=1,
instance_type=instance_type,
)
# run the transform job
transformer.transform(
data=test_data,
content_type="text/csv",
)
After the batch transform job has been completed, we can load the results from S3. For the purpose of evaluating the model’s directional accuracy, we also include the actual and predicted percentage changes in the results.
# get the forecasts from S3
predictions = sagemaker_session.read_s3_file(
bucket=bucket,
key_prefix=f"{transformer.latest_transform_job.name}/test_data.csv.out"
)
# cast the forecasts to data frame
predictions = pd.read_csv(io.StringIO(predictions), dtype=float)
# drop the out-of-sample forecast
predictions = predictions.iloc[:-1]
# add the dates
predictions.index = test_dataset.index
# add the actual values
predictions["y"] = test_dataset["y"]
# add the actual and predicted percentage changes
predictions["r"] = predictions["y"] / predictions["y"].shift(periods=1) - 1
predictions["r_mean"] = predictions["y_mean"] / predictions["y"].shift(periods=1) - 1
# drop the missing values
predictions.dropna(inplace=True)
Actual and predicted EUR/USD daily exchange rate over the test set (from 2024-06-19 to 2024-07-31).
Actual and predicted EUR/USD daily percentage changes over the test set (from 2024-06-19 to 2024-07-31).
Evaluation¶
We evaluate the test set predictions using the following metrics:
The root mean squared error (RMSE) of the predicted values.
The mean absolute error (MAE) of the predicted values.
The accuracy of the signs of the predicted percentage changes.
The F1 score of the signs of the predicted percentage changes.
Performance metrics of predicted EUR/USD daily exchange rate over the test set (from 2024-06-19 to 2024-07-31).
We find that the model achieves a mean absolute error of 0.0012 and a mean directional accuracy of 83.33% on the test set.
We can now delete the model.
# delete the model
transformer.delete_model()
References¶
[1] Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural computation, 9(8), pp. 1735-1780. doi: 10.1162/neco.1997.9.8.1735.
[2] Ayitey Junior, M., Appiahene, P., Appiah, O., & Bombie, C. N. (2023). Forex market forecasting using machine learning: systematic literature review and meta-analysis. Journal of Big Data, 10(1), 9. doi: 10.1186/s40537-022-00676-2.
[3] LeCun, Y., Bottou, L., Orr, G. B., & Müller, K. R. (2002). Efficient backprop. In Neural networks: Tricks of the trade., pp. 9-50, Springer. doi: 10.1007/3-540-49430-8_2.
[4] Yıldırım, D. C., Toroslu, I. H., & Fiore, U. (2021). Forecasting directional movement of Forex data using LSTM with technical and macroeconomic indicators. Financial Innovation, 7, pp. 1-36. doi: 10.1186/s40854-020-00220-2.