Time Series Forecasting in Python a talk by Petr Šimeček
Friday, 14 June, 11:50 in Club
This talk should give you a high-level introduction to time series forecasting in Python. In particular, I want to focus on differences between classical statistical approaches (exponential smoothing, ARIMA…) and more recent machine learning methods (neural networks).
I will also describe the history of M1-M4 time series forecasting competitions and provide examples on public Kaggle datasets.
How is it possible that deep neural networks are superior on one type of data while for a different dataset they give forecast so poor that prof. Makridakis summarized his findings (Makridakis et al., Plos One, 2018) as follows: “The forecasting accuracy of the best ML method was lower than the worst of Stat ones while half the ML methods were less accurate than a random walk.”
In the end, I will provide some hints of which method is likely to work best for your data.
While my background is Probability and Statistics (Ph.D. at MFF UK), I have never truly worked with time series data until two years ago. I still remember many WTF moments in the beginning and hope to share these with the audience.
I have recently moved from California to Brno and joined Central European Ai (CEAi) as an ML Engineer. I keep discovering Python and tweeting about it at @python_tip.