🛠️ Scaler Design

August 29, 2025 · View on GitHub

🧐 What is a Scaler and Why Do We Need It?

A scaler is a class designed to handle the normalization and denormalization of data. In time series analysis, raw data often exhibits significant variations in scale. As a result, models—especially deep learning models—typically do not operate directly on the raw data. Instead, a scaler is used to normalize the data within a specific range, making it more suitable for modeling. When calculating loss functions or evaluation metrics, the data may be denormalized back to its original scale to ensure accurate comparison.

This makes the scaler an essential component in the time series analysis workflow.

👾 How is the Scaler Initialized and When Does It Function?

The scaler is initialized alongside other components within the runner.

For example, a Z-Score scaler reads the raw data and computes the mean and standard deviation based on the training data.

The scaler functions after the data is extracted from the dataset. The data is first normalized by the scaler before being passed to the model for training. After the model processes the data, the scaler denormalizes the output before it is passed to the runner for loss calculation and metric evaluation.

Important


In many time series analysis studies, normalization often occurs during data preprocessing, as was the case in earlier versions of BasicTS. However, this approach is not scalable. Adjustments like changing input/output lengths, applying different normalization ways (e.g., individual normalization for each time series), or altering the training/validation/test split ratios would require re-preprocessing the data. To overcome this limitation, BasicTS adopts an "instant normalization" approach, where data is normalized each time it is extracted.

# in runner
for data in dataloader:
    data = scaler.transform(data)
    forward_return = forward(data)
    forward_return = scaler.inverse_transform(forward_return)

🧑‍🔧 How to Select or Customize a Scaler

BasicTS provides several common scalers, such as the Z-Score scaler and Min-Max scaler. You can easily switch scalers by setting CFG.SCALER.TYPE in the configuration file.

If you need to customize a scaler, you can extend the basicts.scaler.BaseScaler class and implement the transform and inverse_transform methods. Alternatively, you can choose not to extend the class but must still implement these two methods.

🧑‍💻 Explore Further