cancel
Showing results for 
Search instead for 
Did you mean: 

Derivation Window Cannot Exceed Around 150 Days?

Derivation Window Cannot Exceed Around 150 Days?

I downloaded exactly one year's worth of data to get a good view of the Seasonality within my current analyses. That would be around 365 days. However, DataRobot only allows me to use around 150+ days. How should I identify my Seasonality/Calendar with only about a half a year's worth of data?

Labels (1)
0 Kudos
1 Solution

Accepted Solutions
shu
Data Scientist
Data Scientist

Hi Zach,

The derivation window defines the rolling window in creating lagged features: https://docs.datarobot.com/en/docs/data/transform-data/feature-discovery/fd-time.html#prediction-poi... 

The model is not limited to this window to detect seasonality: https://docs.datarobot.com/en/docs/modeling/time/ts-reference/ts-feature-lists.html#intra-month-seas...

So having too long a derivation window is actually the concern that you mentioned for identifying seasonality:

  1. Longer FDW means less training data: with a 150+ day FDW, we'd lose almost half of the 1 year history for the model to learn the trend;
  2. When you select longer FDWs (say, 3 months), DataRobot will automatically derive and test shorter derivation windows (say one week or one month). So 150+ day FDW could take extended time to run.

You can use the API client to build a model factory and test various FDWs. But please align the backtesting periods and forecast windows for an apples to apples comparison. Also, since the data is only 1 year in history, it might be worthwhile to avoid using the month as a feature.

View solution in original post

0 Kudos
2 Replies
shu
Data Scientist
Data Scientist

Hi Zach,

The derivation window defines the rolling window in creating lagged features: https://docs.datarobot.com/en/docs/data/transform-data/feature-discovery/fd-time.html#prediction-poi... 

The model is not limited to this window to detect seasonality: https://docs.datarobot.com/en/docs/modeling/time/ts-reference/ts-feature-lists.html#intra-month-seas...

So having too long a derivation window is actually the concern that you mentioned for identifying seasonality:

  1. Longer FDW means less training data: with a 150+ day FDW, we'd lose almost half of the 1 year history for the model to learn the trend;
  2. When you select longer FDWs (say, 3 months), DataRobot will automatically derive and test shorter derivation windows (say one week or one month). So 150+ day FDW could take extended time to run.

You can use the API client to build a model factory and test various FDWs. But please align the backtesting periods and forecast windows for an apples to apples comparison. Also, since the data is only 1 year in history, it might be worthwhile to avoid using the month as a feature.

0 Kudos

My problem was that I was accepting/thinking the FDW as the amount of data that it would use for the Prediction instead of as the rolling-window. In other words, I was incorrect in saying that if I pulled one year of data to be used as input, it would only use the 150 of the 365 days; not so. The "rolling" window is what it says it is in terms of compartmentalizing the varying runs.

0 Kudos