We live in a world of unknown unknowns: unknown events and unknown measurement changes make it difficult to trust observations as we try to evaluate (and forcast with) complex models we cannot solve analytically. Break detection methods based on indicator saturation can be a useful tool in this world - from detecting unknown events (in my application volcanic eruptions - but also applicable to other schoks), to identifying measurement changes ("Is anyone still using buckets to measure sea surface temperatures?"), and evaluating models where all we have is observations and model predictions ("I cannot solve my model but still want to assess its relative performance"). The techniques developed at INET and presented in this seminar can be useful to anyone working with time series data.
I will provide a whirlwine tour through the world of indicator saturation and how it can be useful (and implemented easily) even if you are more concerned about complexity, sustainability or employment, equity and growth, than volcanic eruptions 800 years ago.