Charting a Course With Only the Wake Behind You
why forecasting is fundamentally different from every other prediction problem: you only ever see where you've already been.
Predicting what comes next when the past is your only guide.
why forecasting is fundamentally different from every other prediction problem: you only ever see where you've already been.
the real difference between describing a sequence's past and genuinely predicting where it's headed next.
why the simplest possible forecast — tomorrow looks like today — is a surprisingly strong baseline that every real model has to beat.
how to separate a sequence's persistent, long-term direction from its short-term noise — the first step in understanding what's actually moving it.
how recognizing regular, repeating cycles in a time series — the tides — separates predictable rhythm from genuine unpredictability.
why every forecast has an irreducible amount of genuine randomness left over, and why that's a fact about the world, not a modeling failure.
how formal decomposition splits a time series into trend, seasonality, and noise as three separate, combinable pieces.
how ARIMA models a time series' own past values and past errors to produce a rigorous, well-understood forecast.
how exponential smoothing weights recent observations more heavily than old ones, adapting gradually as conditions genuinely change.
why forecasting one sequence in isolation misses information that related sequences can provide, and how multivariate forecasting captures it.
how gradient boosting and other machine learning methods reframe forecasting as a standard supervised learning problem with engineered time-based features.
how LSTMs and transformer-based models handle long, complex sequences and multiple related series in ways classical methods can't.
why forecast accuracy degrades predictably the further into the future you look, and why the honest answer to 'how far ahead can we forecast' is 'it depends on the horizon.'
why a single point forecast without an honest uncertainty range is dangerously incomplete, and how prediction intervals fix that.
why a forecasting model trained on historical patterns can quietly stop working once the underlying dynamics genuinely change.
how backtesting lets you rigorously compare forecasting methods on real historical data before trusting any of them with a real decision.
how metrics like MAPE and RMSE turn 'how good was this forecast' into a precise, comparable number, and why the choice of metric genuinely matters.
why combining forecasts from several different methods often beats relying on any single one, however well it backtested.
how hierarchical and large-scale forecasting keep thousands of related predictions coherent with each other, not just individually accurate.
reassembling the whole voyage, from a wake trailing behind an uncertain ship to a coordinated fleet of forecasts reaching landfall together.