This project involved developing an algorithmic trading strategy to help automate the decision-making and execution of intraday trading against the SPY ETF (tracks S&P 500). The solution leverages Alpaca Trading API (for historical data) and the SuperTrend indicator strategy to generate buy/sell signals for intraday trading and to automate order execution. Furthermore, the Amazon DeepAR Forecasting algorithm was integrated as a real-time trading signal confirmation to use in concert with the automated buy/sell execution of the SuperTrend Trading Bot.

The choice for using this indicator was due in part to several reasons. Namely, we wanted an approach that was easy to understand and flexible for many timeframes. Additionally, we liked that it provides precise buy and sell signals and works well with a trending market. Since the scope of our project was time-constrained, we opted to implement a "Long-only" trading strategy.

The Supertrend indicator requires no model training, only the implemetation of the Supertrend Indicator Formula as shown below in figure 1.

According to documentation, the indicator works very well in trending markets. The buy signal is generated with the "Super Trend" closes below the most recent close price. Similarly, the sell signal is generated when the "Super Trend" closes above the most recent close price. Lastly, this trend indicator can be combined with other approaches to generate even "better" signals. That is what we endeavored to find out here...

The DeepAR Forecasting algorithm was selected as it is a supervised learning algorithm for forecasting scalar time series such as stock market closing prices. This, theoretically, will allow for the model to extrapolate the time series data into the future for complementary trading signal confirmation. Moreover, DeepAR outperforms standard ARIMA and Exponential Smoothing techniques when datasets contain hundreds of related time series. Figure 2 depicts is a visual overlay of train and test datasets for a given time series (1 day sample at 15 minute timeframe).

The DeepAR Forecasting model uses various hyperparameters to allow for tuning and optimization of the neural network (RNN). The settings for the parameters that were used for the analysis are shown below:

    epochs=40,
    num_cells=32,
    num_layers=3,
    dropout_rate=0.05,
    mini_batch_size=32,
    learning_rate=0.001,
    likelihood="gaussian",
    time_freq=frequency,
    early_stopping_patience=20,
    context_length=16,
    prediction_length=3

The DeepAR algorithm evaluates the trained model using the root mean squared error (RMSE) over the test data. Figure 3 shows the formulat used for time series data:

Additionally, the algorithm evaluates the accuracy of the forecast distribution using weighted quantile loss. Figure 4 shows the calculations for the weighted quantile loss:

Below is the representative output from training of the model:

    [12/08/2021 21:27:18 INFO 140499396806016] #test_score (algo-1, RMSE): 2.1269036559074226
    [12/08/2021 21:27:18 INFO 140499396806016] #test_score (algo-1, mean_absolute_QuantileLoss): 83.20482245551214
    [12/08/2021 21:27:18 INFO 140499396806016] #test_score (algo-1, mean_wQuantileLoss): 0.0028445597400594325
    [12/08/2021 21:27:18 INFO 140499396806016] #test_score (algo-1, wQuantileLoss[0.1]): 0.0025912611382106652
    [12/08/2021 21:27:18 INFO 140499396806016] #test_score (algo-1, wQuantileLoss[0.2]): 0.003717977404373495
    [12/08/2021 21:27:18 INFO 140499396806016] #test_score (algo-1, wQuantileLoss[0.3]): 0.004196458815721761
    [12/08/2021 21:27:18 INFO 140499396806016] #test_score (algo-1, wQuantileLoss[0.4]): 0.004143057636348619
    [12/08/2021 21:27:18 INFO 140499396806016] #test_score (algo-1, wQuantileLoss[0.5]): 0.003426094011382414
    [12/08/2021 21:27:18 INFO 140499396806016] #test_score (algo-1, wQuantileLoss[0.6]): 0.0026046183189515726
    [12/08/2021 21:27:18 INFO 140499396806016] #test_score (algo-1, wQuantileLoss[0.7]): 0.0019821035437349818
    [12/08/2021 21:27:18 INFO 140499396806016] #test_score (algo-1, wQuantileLoss[0.8]): 0.0015710801414318299
    [12/08/2021 21:27:18 INFO 140499396806016] #test_score (algo-1, wQuantileLoss[0.9]): 0.0013683866503795571
    [12/08/2021 21:27:18 INFO 140499396806016] #quality_metric: host=algo-1, test RMSE <loss>=2.1269036559074226
    [12/08/2021 21:27:18 INFO 140499396806016] #quality_metric: host=algo-1, test mean_wQuantileLoss <loss>=0.0028445597400594325

From the live trading sessions, the Supertrend Indicator strategy worked well in identifying Buy/Sell signals at a 15-minute timeframe. However, because the strategy was "long only", there was one day in our testing session where no trade signals were generated. Overall, we identified that the DeepAR Model did not appear to effectively forecast future prices or our usage of the forecasted values was not appropriate as implemented. Of particular note, the downtrend on 12/09 did not trigger and "buy" signals so feedback was limited. Figure 5 shows the DeepAR Forecast model for end of day which include the closing price targets and corresponding mean predictions with 80% confidence interval for the next 3 successive closing prices. In this case the predictions seem plausible, albeit the confidence interval is quite large and therefore may has questionable utility.

On 12/10, there were several uptrend signals generated from the Supertrend Indicator. However, the DeepAR Forecast was predicting a much lower mean. Given that, and how we implemented the signal confirmation the end result was that the buy signals were repeatedly overriden for the entire day.

Figure 7 provides representative Trading Bot log output which captures the SuperTrend and DeepAR Model Forecast data along with the buy override message.

• Anaconda/Jupyter Notebook
• Python >= 3.7
• Git Bash

• open a git bash terminal on your local computer
• type git clone https://github.com/gakees/project_2_algorithm_trading_strategy.git
• then type cd project_2_algorithm_trading_strategy/Code/
• create a .env file and add the following content:

    ALPACA_API_KEY=[your alpaca api key goes here...]
    ALPACA_SECRET_KEY=[your alpaca secret key goes here...]

    AWS_REGION_NAME=[your hosted region name goes here...]
    AWS_ENDPOINT_NAME=[your sagemaker hosted endpoint name goes here...]

• To run the supertrend_long_trading_strategy_bot.ipynb notebook you will need to...
  • activate the conda algotrading environment and then install boto3, python-dotenv and alpaca-trade-api
  • type conda activate algotrading
  • type pip install boto3[crt]
  • type pip install python-dotenv
  • type pip install alpaca-trade-api
• To run the stock_forecasting_model_setup.ipynb notebook, see the next section below
  • NOTE: When running the SageMaker notebook in AWS, you will need to also upload the .env file into the /Code folder

• login to AWS SageMaker Studio with a suitable role (IAM Access Role preferred)
• Create a new SageMaker Notebook instance and S3 Bucket in the same AWS Region (NOTE: important that they be deployed to same region)
• Within the SageMaker notebook clone the git repository to access the stock_forecasting_model_setup.ipynb notebook
• Run the notebook
  • NOTE: this will take some time for the model training and endpoint deployment
  • When the run is completed the SageMaker endpoint will now be available
  • Go back to the SageMaker Dashboard
    • Access the Inference Menu
      • Select the Endpoints Menu Item
      • This will show a list of Endpoints deployed
      • Make note of the endpoint name and update the .env file (AWS_ENDPOINT_NAME=[your sagemaker hosted endpoint name goes here...])

• from within the cloned git repository
  • type jupyter lab (from git bash)
  • then open the supertrend_long_trading_strategy_bot.ipynb file within the browser (/Code/supertrend_long_trading_strategy_bot.ipynb)