My personal journey developing a plugin for Dataiku DSS to add visual recipes to solve time series forecasting problems. This plugin offers a set of 3 visual recipes to forecast yearly to hourly time series. It covers the full cycle of data cleaning, model training, evaluation and prediction including multivariate forecasting and partitioned data.
2. What is a time series?
2
@alex_combessie | @dataiku | @datasciencefest
3. What is (time series) forecasting?
3
@alex_combessie | @dataiku | @datasciencefest
4. Why is it needed?
Predict a Price
(Nintendo)
Plan an Economy
(Japanese GDP)
Predict Sales/Demand
(Visits to Japan)
Finance
Public
Retail / CPG
/ Travel
4
@alex_combessie | @dataiku | @datasciencefest
5. 一 Just ファック-ing Demo.
一 Origin Story
二 Forecasting VS Machine Learning
一 Choose Your Stack
二 Evaluate Your Performance
三 Make It Scale
2. Demo
1. Why?
3. How?
Agenda
5
@alex_combessie | @dataiku | @datasciencefest
6. 一 Origin Story
二 Forecasting VS Machine Learning
1. Why?
6
@alex_combessie | @dataiku | @datasciencefest
8. Wait, we have a 20% time rule?
2+ year talking about it.
40+ man-days to actually just do it.
Origin Story
A 20% Project at Dataiku
8
@alex_combessie | @dataiku | @datasciencefest
9. Forecast VS Machine Learning
Specificities of Forecasting
Multiple output instead of one
Evaluation strategy needs to respect time
Specific algorithm implementations |
9
@alex_combessie | @dataiku | @datasciencefest
14. Choose Your Stack
A Little Bit of Nuance
14
Don’t Be a Dick Dictator
Forecast + Prophet VS statsmodel:
R wins!
Not everything is Open Source 😢
@alex_combessie | @dataiku | @datasciencefest