Last week we were fortunate to have had one of the foremost statistical forecasters in the world, Dr. Rob Hyndman, join us for a Big Thinkers talk on the science of forecasting and prediction. Dr. Hyndman delivered a captivating seminar called, “Exploring the boundaries of predictability: what can we forecast, and when should we give up?”
In his talk, Dr. Hyndman discussed his research efforts to realize his long-term goal of understanding what models work best at predicting various types of time-series. Professor Hyndman demonstrated the aspects that make certain time-series more predictable than others by giving real-life examples based on his own research and work consulting for the Australian government. He explored different features that can be extracted from time-series and how they can be used to judge if something is predictable. Professor Hyndman also talked about his work with Yahoo Labs on Anomaly Detection and how extracted time-series features can be used for finding anomalies. While Dr. Hydnaman’s goal of finding the right time-series model for a given dataset is still under active pursuit, his work on anomaly detection in collaboration with Yahoo Labs has already been open-sourced.
If you would like to learn what it takes to forecast and accurately predict the future, watch the full seminar here.
Great overview of basic forecasting methodology from Rob Hyndman.
Rule of thumb – ok to forecast ahead a third of the length of your data. So, if you have 15 years of data, a five-year forecast should be pretty reliable. But doing a 20 year forecast based on 15 years of data is going to be pretty unreliable.