CODEWITHSUNDEEP
h2oforecastingdriverless-ai
spokes
spokes

Reputation: 1

Forecasting with a gap of negative 1

I am using H2O.ai to understand both current week and lagged week features that affect the target value of the current week.
Using the Walmart example.

For a particular week of sales, I am interested in the features that most likely affect how well or poorly sales performed. To set this problem up, I want H2O.ai to 'predict' what the current week of sales are using the current week feature values as well as the lagged values (including the lag of the target) -- i.e., this is not a forecast problem, but a problem to understand the drivers.

In forecasting terms, this would be setting the prediction horizon to 1 and the gap of negative 1.

However, H2O.ai seems to not allow you to set it in this unconventional way.

How would I set up this experiment in H2O.ai?

Upvotes: 0

Views: 56

Answers (1)

topchef
topchef

Reputation: 19793

The problem described corresponds to a multi-variate time series problem that uses only lags 0 and 1. Of course, it can't and won't use use lag 0 on target, but it can will use lag 0 on other time series and features.

To setup the problem in Driverless AI properly with walmart dataset set:

  • target column: Weekly_Sales
  • time column: Date
  • time group columns (TGC): Store, Dept
  • Forecast horizon: 1 week
  • Gap: 0 week
  • Expert Settings -> Time Series:
    • Time Series Lag Override: 1
    • (optional but recommended) Probability to create non-target lag features: 0.5 or higher
    • (optional but worth trying) Always Group by All Time Groups Columns for Creating Lag Features: Disabled

With lag override set to 1 no lags greater than 1 can be used to predict target, plus current (lag 0) features (everything but target). So, effectively, it still a forecast problem but it complies with all restrictions you placed in your question.

Upvotes: 0

Related Questions