How we trained our base MLS model

Here's a guide explaining how the base MLS model was trained:

Github Repo > https://github.com/sportstensor/MLS

1. Data Retrieval and Preparation

The process begins with the get_data function in retrieve_data.py. This function is responsible for collecting and organizing MLS fixture data:

• It scrapes data from FBRef for the past 5 years of MLS matches (excluding 2020).

• For each match, it collects:

  • Date

  • Home Team

  • Away Team

  • Home Team Goal Difference (before the match)

  • Away Team Goal Difference (before the match)

  • Home Team Score

  • Away Team Score

• The data is stored in an Excel file (mls_fixture_data.xlsx).

2. Data Scaling

The scale_data function in retrieve_data.py prepares the data for model input:

• It scales the input features (team IDs, goal differences) and target variables (scores) using MinMaxScaler.

• This normalization helps in improving model performance.

3. Model Architecture

The model is defined in the load_or_run_model function in model.py:

• It's a simple Sequential model using Keras.

• The model consists of an input layer and a single dense layer with 2 units and ReLU activation.

• The model uses Adam optimizer and Mean Squared Error as the loss function.

4. Training Process

The training process, also in load_or_run_model, includes:

• Splitting the data into training and test sets (80% train, 20% test).

• Using Early Stopping to prevent overfitting, with a patience of 6 epochs.

• Training for a maximum of 150 epochs with a batch size of 32.

• Saving the best model based on the lowest loss.

5. Model Evaluation

After training, the model is evaluated on the test set:

• It calculates the percentage of correct score predictions and correct outcome predictions.

• It computes Mean Squared Error, Mean Absolute Error, and R-squared values for both home and away score predictions.

6. Model Usage

The trained model is used in the activate function in miner_dashboard.py:

• It loads the trained model and scalers.

• For given fixtures, it prepares the input data using the prep_pred_input function.

• It then uses the model to predict scores for these fixtures.

• The predicted scores are inverse-transformed to the original scale.

7. Additional Notes

• The model uses a unique numerical encoding for each team, based on their historical presence in the MLS.

• The data collection process uses proxy rotation to avoid IP blocks during web scraping.

• The model can handle both historical and future fixtures, adjusting its input preparation accordingly.

This base model provides a foundation for predicting MLS match outcomes, which can be further refined and improved in future iterations.