MIW 5

What are the assumptions of the autoregressive (AR) model and what limitations arise from them?

An AR model assumes a stationary linear relationship between current and past values. Its limitation is that financial data is often nonlinear, noisy, non-stationary, and affected by external events, so AR models may not generalize well.

How does the choice of the AR model order affect its generalization capability?

A low AR order may underfit because it ignores useful past information. A high order may overfit noise. Good generalization requires choosing an order that captures real dependencies without memorizing the training data.

What are the differences between LSTM and GRU architectures, and in which scenarios might each be more advantageous?

LSTM is more complex and uses separate memory with input, forget, and output gates, so it can model long dependencies well. GRU is simpler, faster, and has fewer parameters, so it is often better when data or computation is limited.

Why is input data normalization important when training RNN models?

Normalization puts all inputs on a similar scale, which makes RNN training more stable and faster. It prevents large-valued features like volume from dominating smaller features like returns.

What is the role of additional features (e.g., technical indicators, volume) in time series modeling?

Additional features can provide information about trend, momentum, volatility, and market activity. But they are only useful if they improve test performance; otherwise they may add noise and cause overfitting.

How should MSE and MAE metrics be interpreted, and what are their advantages and disadvantages in financial prediction tasks?

MAE shows the average absolute error in the same unit as the target, so it is easy to interpret. MSE squares errors, so it penalizes large mistakes more strongly, but it is less interpretable and more sensitive to outliers.