Tutorial 3: How to Develop Strategies Using Machine Learning Models

1. Training Data

In this step, you will define the specifications of the training data. In other words, you will select the dataset that the model will use during the training process. To proceed to the next step, the following fields must be completed:

• Name: Assign a meaningful name to your model to help identify it easily.

• Market Data: Select the market data that you previously created for training.

• Training Pair(s): Select which pairs you want to use for training. In this tutorial, we will use all the pairs included in the selected market data.

• Timeframe: If your market data contains multiple timeframes, select the specific timeframe you want the model to be trained on.

• Numeric Features: Choose the features that the model will use for training. You can select multiple features, but keep in mind that including too many may increase the training time. In this tutorial, we will use OHLCV, RSI, and Bollinger Bands.

• Target: Define the target variable that the model will learn to predict. The model will be trained to mimic the behavior of the target as closely as possible. In other words, the goal of training phase is to predict the target labels by using the selected features. In this tutorial, we will use Rally-Based Classifying with window size of 100, buy and sell thresholds of 0.2.

• Start Date: Specify the starting date for the training data.

• End Date: Define the last date of the dataset to be used for training.

Note that a fair and reliable evaluation of a model is only possible when the training period and test period are kept separate. This means that the model must be tested on a time period that is not included in the training phase.

2. Model Parameters

As the name suggests, this step involves configuring the parameters of the model, which serves as the core of the machine learning process. Our platform currently offers five model options:

• XGBoost Classifier: A tree-based model that leverages the power of eXtreme Gradient Boosting (XGBoost) to enhance predictive performance.

• AdaBoost Classifier: A tree-based model that employs adaptive Boosting.

• K-Nearest Neighbor (KNN): A non-parametric classifier that makes predictions based on the closest previous samples to the current test sample.

• Naïve Bayes: A probabilistic classifier that applies a simplified version of Bayes' Theorem.

• CatBoost: A tree-based model similar to XGBoost, but optimized for lower computational cost and featuring structural differences that enhance efficiency.

Each model has its own advantages and disadvantages—there is no single model that always delivers the best performance. It is up to you to experiment and determine which model works best for your dataset.
Additionally, each model comes with various parameters that influence its behavior. You can access these configurations by clicking on Advanced.
To keep things simple, we will use the XGBoost Classifier with its default settings in this tutorial. You are encouraged to modify the parameters and observe how they impact the results. However, fine-tuning models is beyond the scope of this tutorial. For a deeper understanding of the available models and their parameters, refer to the documentation.

3. Feature Processor

The Feature Processor plays a crucial role in preparing data for the model. It serves two main purposes:

• Shaping the input of the model

• Pre-processing the inputs of the model

Model Memory defines the number of historical candles the model considers for inference.

• When memory is set to 1, the model only uses the features of the last candle to predict the label for the candle.

• As the memory size increases, the model incorporates data from more previous candles, potentially improving performance by recognizing longer-term patterns.

Note that increasing memory also leads to longer training times and does not always guarantee better results. As an analyst, it is your responsibility to experiment and determine the optimal configuration for your model.
In this tutorial, we will use a memory size of 1 for simplicity.
If you open the Advanced dropdown menu, you will find additional options for sampling and input data processing. For this tutorial, we will keep the default settings, but if you’re interested in learning more, refer to the documentation.

Price-Like Feature Normalization is a process that normalizes features measured in dollar values ($). This transformation helps the model learn patterns in price data more effectively—though its effectiveness may vary depending on the dataset.
One consequence of this normalization is that the values of one feature may become zero for all candles. However, this adjustment generally contributes to a more efficient learning process.
For this tutorial, we will not modify these settings. If you want to explore this topic further, refer to the documentation.

Once you have configured the Feature Processor, the model is ready for training.
To begin, click on the Train button. After doing so, you will be redirected to the My Models page. Wait for the status of the model to become Successful.
Note: Training process may take several minutes to complete, especially when using large memory sizes or a high number of features.
Once the model is successfully trained, you can use the Actions menu to:

• Delete the model

• Clone the model

• Statistically evaluate its performance.

We will explore these functionalities in the next tutorials.