Research first forecasting approach for market time series

[BytecodeBrewer]

Goal

Research a first simple forecasting approach for ARGUS market time-series data.

Why

ARGUS should eventually support basic prediction and signal workflows. Before implementing forecasting, the project should define a realistic first approach and avoid jumping too early into complex deep learning models.

This ticket should compare simple baseline methods with first machine-learning options and clarify what data, metrics and evaluation are needed.

Research questions

• What is a realistic first prediction task for ARGUS?
  • next-day exchange-rate movement
  • next value prediction
  • trend direction
  • volatility expectation
• What baseline methods should be implemented first?
  • naive last-value forecast
  • moving average forecast
  • simple linear regression
• Should the first implementation use NumPy, pandas or scikit-learn?
• What evaluation metrics should be used?
  • MAE
  • RMSE
  • directional accuracy
• Why is LSTM not the first implementation step?
• What would be required before an LSTM ticket becomes realistic?

Acceptance criteria

• First prediction task is defined
• Simple baseline approaches are compared
• scikit-learn / NumPy / pandas usage is evaluated
• Evaluation metrics are documented
• LSTM requirements are documented as future work
• Recommended first implementation approach is selected
• No production prediction model is required in this ticket

Note

Priority: Should