Nowcasting Business Performance

talk by Francois Bouet at Data Science London, 28/11/12

Transcript

1. NOWCASTING BUSINESS PERFORMANCE miércoles, 28 de noviembre de 12
2. GROWTH INTELLIGENCE What we do Classification of companies Revenue estimation What we use Machine Learning Times Series methods miércoles, 28 de noviembre de 12
3. SOME OF OUR CLIENTS miércoles, 28 de noviembre de 12
4. NOWCASTING Estimating the value of a time series not readily available at present present miércoles, 28 de noviembre de 12
5. NOWCASTING present miércoles, 28 de noviembre de 12
6. NOWCASTING Previously called short-term forecasting forecasting More an approach and a goal than a different theory and field miércoles, 28 de noviembre de 12
7. NOWCASTING USE CASES Weather nowcasting Search-based nowcasting GDP nowcasting miércoles, 28 de noviembre de 12
8. WEATHER NOWCASTING Simplified model that is applied quickly Uses weather models Forecast at location x given weather at y → Not applicable to other fields miércoles, 28 de noviembre de 12
9. SEARCH-BASED NOWCASTING Popularized by Google Recent successes Flu predictions Consumer behaviour travel, movies and products Based on Google’s data, simple AR models Only used to study what people are searching for miércoles, 28 de noviembre de 12
10. miércoles, 28 de noviembre de 12
11. miércoles, 28 de noviembre de 12
12. GDP NOWCASTING Field with the most generic research Major research since the 90's GDP released quarterly with further revisions 1000's of signals for GDP nowcasting Industrial production, unemployment, confidence surveys, retail sales, ... miércoles, 28 de noviembre de 12
13. GDP NOWCASTING Vector auto-regression and the “jagged edge” Present Different frequencies, different lag, missing data miércoles, 28 de noviembre de 12
14. miércoles, 28 de noviembre de 12
15. miércoles, 28 de noviembre de 12
16. miércoles, 28 de noviembre de 12
17. Search results LinkedIn info Web traffic Patents Advertisement spending Liabilities Tweets Assets Website updates Press miércoles, 28 de noviembre de 12
18. TIE WITH “BIG DATA” Need to gather signals in large quantity Machine learning as a pre-processing step and to integrate discrete events Example: companies in a sector which receive investment miércoles, 28 de noviembre de 12
19. TIE WITH ESTIMATION THEORY Beneath all this: Getting to a variable not directly observable with the help of measured signals Replacing probability distribution from physical models with machine learned knowledge miércoles, 28 de noviembre de 12
20. METHODOLOGIES Vector auto-regression Challenge with large number of signals (predictors): Curse of dimensionality when applying VAR Machine Learning approach Own solution: ziggurat miércoles, 28 de noviembre de 12
21. TIME SERIES + MACHINE LEARNING Δrevenue avg, std dev, model params miércoles, 28 de noviembre de 12
22. OUR PIPELINE FOR NOWCASTING Clustering companies in sets (ML) Signals gathering Time Series processing ML with model for each cluster > Revenue for each company and each cluster miércoles, 28 de noviembre de 12
23. TECHNOLOGIES miércoles, 28 de noviembre de 12
24. DATA SCIENCE AT GROWTH INTELLIGENCE :D miércoles, 28 de noviembre de 12