Forecasting Based on Time series Historical Data

Forecasting Based on Time series Historical Data

Transcript

1. Disclaimer Presentations are intended for educational purposes only and not to replace independent professional judgment. Statements of facts, views, opinions expressed do not necessarily reflect those of Blibli.com.
2. Forecasting based on Time Series Historical Data
3. Profile • Uriando Tjiangdiono • R&D Engineer at Blibli.com • Part of R&D Team at Blibli.com • uriando.tjiangdiono@gdn- commerce.com
4. Outline • Forecasting • Historical Data • Time series • Prophet • Make a Forecast • Question
5. Forecasting
6. Forecasting • fore·cast \ ˈfȯr-ˌkast ; fȯr-ˈkast \ • to calculate or predict (some future event or condition) usually as a result of study and analysis of available pertinent data
7. Historical Data • Past-periods data, used usually as a basis for forecasting the future data or trends – 1 Year ago – 1 Month ago – 1 Day ago – 1 Second ago Read more: http://www.businessdictionary.com/definition/his torical-data.html
8. Time Series • a set of data collected sequentially usually at fixed intervals of time (Period)
9. Time Series Component
10. Time Series Component
11. Time Series Component • The trend is the long term pattern of a time series. Trend (T) • Any pattern showing an up and down movement around a given trend is identified as a cyclical pattern Cyclical (C) • Seasonality occurs when the time series exhibits regular fluctuations during the same month (or months) every year, or during the same quarter every year. Seasonal (S) • This component is unpredictable. Every time series has some unpredictable component that makes it a random variable Irregular/Random (I)
12. Use Cases • Sales Growth • Web Traffic • Stock Price • Stock and Demand Purchase • Human Population • ground Temperature • Capacity planning to efficiently allocate resources • etc
13. Prophet • Forecasting at scale
14. Prophet • Procedure for forecasting time series data • Open source software released by Facebook’s Core Data Science team • Implemented in Python and R. • Provides completely automated forecasts that tunable • https://facebook.github.io/prophet
15. Motivation
16. Motivation
17. Motivation • Producing high quality forecasts is not an easy problem • Completely automatic forecasting techniques can be brittle and often too inflexible. • Analysts who can produce high quality forecasts are quite rare.
18. Workflow
19. Prophet Accurate and Fast Fully Automatic Tunable forecasts Available in Python and R
20. Where Prophet Shine • Hourly, daily, or weekly observations with at least a few months (preferably a year) of history • Strong multiple “human-scale” seasonalities: day of week and time of year • Important holidays that occur at irregular intervals that are known in advance (e.g. Harbolnas) • Missing observations or Large outliers • Shift in the trend • Trends that are non-linear growth curves, where a trend hits a natural limit or saturates
21. Message from Creator • “We have found Prophet’s default settings to produce forecasts that are often accurate as those produced by skilled forecasters, with much less effort.” • “With Prophet, you are not stuck with the results of a completely automatic procedure if the forecast is not satisfactory — an analyst with no training in time series methods can improve or tweak forecasts using a variety of easily-interpretable parameters.”
22. Example • Number of Bicycle Hires in London TFL(Transport for London) From 2010 - 2017 • https://data.london.gov.uk/dataset/number- bicycle-hires
23. Example
24. Example
25. Example
26. How It Works y(t)=g(t)+s(t)+h(t)+ϵt Where • y(t) is the variable we are predicting • g(t) models the growth (trend) of the data • s(t) models the periodic (seasonal) component of the data • h(t) is models the impact the holidays have on the data, • ϵt is the model error at time t
27. Install • pip Install fbprophet • conda install -c conda-forge fbprophet • Requirements – Cython>=0.22 – pystan>=2.14 – numpy>=1.10.0 – pandas>=0.18.1 – matplotlib>=2.0.0
28. Input a dataframe (pandas) with two columns: • ds = Datestamp • y = must be numeric and represents the measurement we wish to forecast.
29. Main Feature • Saturating Forecasts • Trend Changepoints • Uncertainty Intervals • Seasonality And Holiday Effects • Non-Daily Data • Diagnostics
30. Demo Demo
31. Advantages • It’s fast, really fast! • Handles daily, weekly, monthly or yearly. • Completely automated forecasts that tunable • Irregular holidays can be taken care of.
32. Advantages • In case the forecast is going beyond a certain limit based on business understanding, it can be fixed by setting up a forecasting cap and modelling using logarithmic growth instead of linear growth. • Outliers can be handled well by model itself without any requirement for imputation.
33. Limitations • In case your time-series is highly irregular • Exogenous variables affecting the time series cannot be taken care using prophet. • Multiplicative models cannot be accounted for using prophet.
34. Conclusion • Producing high quality forecasts is not an easy problem • Prophet fills in this gap and provides a very intuitive forecast and easy to use prophet function that can be used for forecasting
35. References • https://peerj.com/preprints/3190.pdf • https://facebook.github.io/prophet • https://github.com/facebook/prophet
36. THANK YOU Any question?
37. We Are Hiring