Soothsayer: Using Neuralprophet, Axon, Nx and Livebook to Forecast Business Data in Elixir

Talk I gave at ElixirConf US 2024 in Orlando, Florida.

Transcript

1. USING NEURALPROPHET, AXON, NX AND LIVEBOOK TO FORECAST BUSINESS DATA IN ELIXIR George Guimarães Principal Engineer at CloudWalk Elixir Advocate
2. YEAH, LET’S FORECAST STUFF USING ALGORITHMS! Nothing to worry about, right?
4. LEARN TO LOSE $300M FAST USING ELIXIR AND FRIENDS!
5. GEORGE GUIMARÃES Principal Engineer at CloudWalk Previously - Co-Founder at SourceLevel - Co-Founder at Plataformatec
8. PROPHET • Facebook Project • Predicts time series, paper from 2017 • Simple to use • Written using Stan, a statistical modeling platform and language • https://github.com/tmbb/ulam_ex
9. NEURALPROPHET • Stanford project, paper from 2021 • Rewritten using neural networks (PyTorch) • As with Prophet, focused on “understability”, “explainability” • breaks the prediction into reasonable components • trend, seasonalities, holidays/events • easy for a human to review • Simple to use
14. Linear trend (offset and rate of change). It can change over time. TREND
15. Periodic movement of the season. Modeled using Fourier terms. SEASONALITY
16. AUTO REGRESSOR Use a variable past values to predict its future values.
17. LAGGED REGRESSION Effect of variables known in the past. Modeled very similarly to Auto Regression. • Other variables in your “vicinity" that may be good predictors of behaviours • sales for A/C units: temperature
18. FUTURE REGRESSION Effect of variables known in the future.
19. HOLIDAY EFFECT Occasional events that may cause effects.
21. • SourceLevel, startup doing Engineering Ops, spin-off from Plataformatec • time-series data to predict and/or analise • Engineering Management metrics like Pull Requests, Lead Times, “activity” levels for individuals and teams • Great tools in Python, but no tools in Elixir • So…
22. SOOTHSAYER Introducing
24. kinda, no piecewise linear, and no custom seasons yet
25. LIVEBOOK AND CODE TIME
27. ROADMAP • Fill in missing data • Add Fourier series for seasons TBD Done Preprocessor • Piecewise Linear Trend • Multiplicative Seasonality • Custom seasons (monthly, for instance) • Auto regressor • Lagged and future regressors • Adding holidays/events • Global Linear Trend • Additive Seasonality Components • Make it easier to generate component output • Make it easier to use Nx.serving • Use the same testing data as NeuralProphet General
28. georgeguimaraes/soothsayer georgeguimaraes/awesome- machine-learning-elixir
29. THANK YOU @georgeguimaraes georgeguimaraes/soothsayer