[ENH] Auto naive forecaster #2926
Conversation
aeon-actions-bot
bot
added
the
enhancement
Thank you for contributing to
|
| return self._fitted_scalar_value | ||
| # For "seasonal_last" strategy | ||
| prediction_index = (self.horizon - 1) % self.season_ | ||
| return self.self._y [prediction_index] |
There was a problem hiding this comment.
I believe this should be self.y[-self.season - 1 + prediction_index)] that way it is predicting the last seasonal value in the series, rather than the corresponding seasonal value at the start of the series
| Naive forecaster with strategy set based on minimising error. | ||
|
|
||
| Searches options, "last", "mean", and "seasonal_last", with season in range [2,max_season]. | ||
| If max_season is not passed to the constructor, it will be set to series length/2. |
| max_season = len(y)/4 | ||
| best_s = None | ||
| best_seasonal = np.inf | ||
| for s in range(1, max_season + 1): |
There was a problem hiding this comment.
Should be 2 surely? - as the interval is [2, max_season]
| Searches options, "last", "mean", and "seasonal_last", with season in range [2,max_season]. | ||
| If max_season is not passed to the constructor, it will be set to series length/2. | ||
|
|
||
| Simple first implementation, splits the train series into 70% train and 30% validation split |
There was a problem hiding this comment.
I don't believe it does this, it simply minimises RMSE over the whole series passed in _fit
Simple forecaster that chooses either naive, mean or seasonal based on train error (or "in sample" as they call it).
Seems like a sensible benchmark to me, although Ive not seen it used. Main issues are
could easily cross validate, but seems OTT for this algorithm.