Forecasting on Ryan Orban

sktime — Unified Machine Learning with Time Series

me@ryanorban.com (Ryan Orban) — Mon, 21 Sep 2020 00:00:00 +0000

Summary

sktime is an open-source Python library from the Alan Turing Institute that provides a unified, scikit-learn-compatible API for time series machine learning. The ecosystem problem it solves: time series tasks (forecasting, classification, clustering, anomaly detection) have historically required different libraries with incompatible interfaces — statsmodels for classical forecasting, tslearn for classification, custom code for everything else. sktime unifies these under a single framework.

Surfing Silver: Dynamic Bayesian Forecasting for Fun and Profit

me@ryanorban.com (Ryan Orban) — Thu, 14 Apr 2016 00:00:00 +0000

Summary

These slides from the Data Popup conference cover dynamic Bayesian forecasting — applying Bayesian state-space models to time series prediction. The dynamic in the title means the model parameters are allowed to evolve over time rather than being fixed, which makes it appropriate for non-stationary series: sales with seasonality shifts, user behavior that changes gradually, demand forecasting where the underlying patterns drift.

Weather Forecasting with Twitter and Pandas

me@ryanorban.com (Ryan Orban) — Mon, 16 Dec 2013 00:00:00 +0000

Summary

This ŷhat blog post (ŷhat was an early model deployment platform for data scientists) demonstrates using Twitter emoticon frequency as a proxy for weather conditions. The hypothesis: on sunny days people tweet more positive emoticons, on rainy days more negative ones. Validate this against a Kaggle weather dataset using pandas for data wrangling and analysis.