Multistep forecast models?

[caseytomlin]

Is it the case that all models provide strictly recursive/autoregressed forecasts? If so, is it on the roadmap to support "direct" multistep models, like as wrappers around any sklearn models that support vector output? If this is already supported, could someone point me to a usage example? Thanks!

[jmoralez]

Hi @caseytomlin, can you provide an example of what you mean? Do these models take the same input and produce the predictions for the whole window? If that is the case you could use mlforecast to do the feature engineering with MLForecast.preprocess and then fit those models to the output.

[jmoralez]

I believe those fit a model to each horizon, right? So it'd be using a model for each horizon instead of the recursive strategy? We do have that on the roadmap.

[jmoralez]

We've added that approach in #80. Please let us know if that fits your use case.

[caseytomlin]

Thanks, @jmoralez ! This is a great addition, though I actually had something different in mind. There are three general methods I can think of:

1. Recursive: train one 1-step model, predict max_horizon = n by recursively appending predictions as input (already implemented)
2. Multi-model: implemented in #80
3. Single model, multi-output: trained on examples (X, y) where each y has length max_horizon

So only 3. is missing as far as I understand.

[jmoralez]

Do you have an example model that does that? xgboost does 2. internally

[caseytomlin]

ah ok, this was a misunderstanding on my part - apologies for the confusion and thanks for the clarification. It seems what I had in mind is still in active development.

[DAT-FYAYC]

To me the documentation of recursive strategy is unclear. So is the documentation for Multi-output model.
As stated in the documentation/code: "By default mlforecast uses the recursive strategy, i.e. a model is trained to predict the next value and if we're predicting several values we do it one at a time and then use the model's predictions as the new target, recompute the features and predict the next step."
How exactly is recomputing of features done? As far as my use case shows the features are simply duplicated from the previous time step.

How are "dynamic_dfs" in the "predict" method handelt in general as well as related to recursive and multi-model?

[DAT-FYAYC]

@jmoralez Thank you for the explanation.

Regarding LightGBMCV this makes sense to me.

Concerning predictions this is still unclear to me. If I have an instance of MLForecast object (regardless if it was trained via the fit method or via cv) and I want to do a prediction for new (unseen) data. I need to provide the input properly. Depending on the strategy the input should vary when using dynamic features. Since with the recursive strategy e.g. for a forecasting horizon=2, the dynamic features of 2 time steps must be provided. Which is not the case for the multi-model strategy. As described above. I tried to check the code but find it hard to verify my understanding.

Would appreciate it a lot if you could confirm or correct my understanding.

[jmoralez]

That's correct. Since the recursive strategy computes predictions one at a time you need to provide the dynamic features for the next horizon timesteps, whereas for the multi-model you only need to provide the values for the next timestep, because that approach computes all the horizon in one go. If you provide more they will be ignored though, so you should be able to use both approaches passing the dynamic dfs for all the horizon.

[DAT-FYAYC]

Great, thanks a lot for the clarification.

[jmoralez]

We'd welcome a PR that clarifies this in the docs if you have time. Also please open an issue if you run into any problems.

[DAT-FYAYC]

@jmoralez I am happy to create a PR that clarifies this in the docs once I tested it.