Forecasting API v2 - events

Published on by Joannes Vermorel.

We provide a Forecasting API that lets client applications (either provided by Lokad or by other people) upload their data and download their forecasts.

The current Forecasting API has been virtually left untouched since the Lokad launch early 2007, but we are currently developing the Forecasting API v2 and we are planning for a release late November.

Note: The Forecasting API v1 will be maintained, the v2 will just be available side-by-side next to the v1.

The main limitation of the API v1 is the lack of time-series meta-data. Indeed, so far, Lokad has been providing a pure time-series framework where only list of time-values could be send to Lokad.

As people have been pointing out, this framework does not make possible to add information on top of the time-series.

For example, if you have a call center, and the call center is down for 2h because of a power outage, you want to be able to tell to Lokad: we had zero calls during those 2h because of a power outage, not because nobody called.

Lokad was treating such events as noise - yet, with extra-information, it becomes possible to properly handle events, instead of ignoring them (at best).

For example, for call centers, typical events include marketing operations, such as mail advertising or service outages. For retailers, it would be product promotions or inventory shortage. And more specifically for eCommerce, it would be front page product display and product inclusions in newsletters.

More technically, Forecasting API v2 introduces the notion of events where an event has several properties:

  • a date of start
  • a duration
  • a name

All events having the same name are supposed to reflect the same phenomenon. For example PowerOutage could be used to name all the events related to an actual power outage of the call center.

By analyzing the impact of events on the actual time-series, Lokad will be able to refine further its forecasts.

But this is not all about the API v2, there is more. Next time we will discuss about forecasting time-series when there is no historical data. Stay tuned.

Categories: forecasting, insights, roadmap, web services Tags: