Forecasting by Relaxation

an adaptive econometric ensemble · real FRED data

Real US retail sales (FRED RSXFSN, 1992–2026). Six forecasting models each predict 6 months ahead; none is best everywhere. Relaxation labeling picks the blend — prior = each model's recent accuracy, compatibility = temporal coherence (no flip-flopping), noise = the shocks nobody forecasts. The gold ensemble tracks the real series, and the strip below shows which model the engine trusts, month by month.

actualHRL ensemble (6-mo-ahead)

What's real & honest. The data is the actual FRED retail-sales series. The six models (naive, seasonal-naive, seasonal-drift, linear-trend, drift, long-mean) are computed causally — each forecast at month t uses only data through t, and the prior weights use only errors already realized (no peeking ahead). The noise band swells around 2020: the engine honestly says "no model fits this," rather than pretending. The scoreboard below is the result: the relaxation ensemble's mean error beats every single model and the plain average — because it weights them by context.

← Cord of Three Strands · The Four Poles · Genesis · data: FRED RSXFSN