Ridge Regression Forecasting with Attention-based Context Vectors
contextridge2f.RdThis function performs ridge regression forecasting using attention-based context vectors as external regressors. Context vectors are computed from the training data using various attention mechanisms and then used to enhance the forecast model.
contextridge2f(
y,
h = 5L,
attention_type = "exponential",
window_size = 3,
decay_factor = 5,
temperature = 1,
sigma = 1,
sensitivity = 1,
alpha = 0.5,
beta = 0.5,
...
)Arguments
- y
A multivariate time series object.
- h
Integer. Forecast horizon. Defaults to 5.
- attention_type
String specifying the type of attention mechanism. Options are: "cosine", "exponential", "dot_product", "scaled_dot_product", "gaussian", "linear", "value_based", "hybrid", "parametric". Default is "exponential".
- window_size
Integer parameter for window size (applicable for "cosine" attention). Defaults to 3.
- decay_factor
Double for decay factor (applicable for "exponential" and "hybrid" attention). Defaults to 5.0.
- temperature
Double for temperature (applicable for "scaled_dot_product" attention). Defaults to 1.0.
- sigma
Double for sigma (applicable for "gaussian" attention). Defaults to 1.0.
- sensitivity
Double for sensitivity (applicable for "value_based" or "hybrid" attention). Defaults to 1.0.
- alpha
Double for alpha (applicable for "parametric" attention). Defaults to 0.5.
- beta
Double for beta (applicable for "parametric" attention). Defaults to 0.5.
- ...
Additional arguments passed to
ahead::ridge2f(e.g., lags, lambda_1, lambda_2, nb_hidden, etc.).
Value
An object returned by ahead::ridge2f for the forecast,
typically a list including mean and prediction intervals.
Details
This approach allows the model to leverage temporal dependencies captured by attention mechanisms, potentially improving forecast accuracy by incorporating weighted historical information.
