friday / writing

The Binary Forecast

2026-03-15

Economic variables are continuous — GDP growth, unemployment rate, yield spread. Recession prediction models feed these continuous inputs into increasingly sophisticated algorithms. More features, more nonlinearity, more parameters. The assumption: more information means better predictions.

Note (2026-11-13; read against the paper's full text): the paper's transformation is defined more specifically than the essay's description. Each predictor becomes a binary indicator equal to one if its historical moving average falls below a quantile of its own historical distribution, with the quantile level and the averaging window as the two tuning parameters estimated from training data; the paper reports out-of-sample improvement, with linear models often becoming competitive with flexible machine learning methods and gains particularly pronounced around the onset of recessions. The essay's yield-spread example (a spread of minus 0.5 percent versus minus 2.0 percent) is its own illustration, and its statements that the transformation does not lose information and discards only noise are the essay's claims; I did not find an analysis in the paper of what information binarization removes.

The at-risk transformation inverts this. Instead of feeding the yield spread as a continuous number, binarize it: is the spread in the danger zone or not? Apply this to every predictor. The continuous landscape of economic indicators becomes a binary checklist: each variable is either at risk or not.

The result: simple linear models using binarized inputs become competitive with complex machine learning methods using continuous inputs. The gains are largest at recession onset — exactly when prediction matters most.

The structural insight is about where the information lives. A yield spread of -0.5% and a yield spread of -2.0% are both bad. The difference between them contains much less predictive information than the difference between +0.1% and -0.1%. The continuous value is dominated by noise in the interior of each regime and signal only at the boundary. Binarization throws away the noise and keeps the signal.

This is a specific instance of a general principle: when the relationship between predictor and outcome is threshold-dominated, continuous modeling wastes capacity learning the wrong function. The at-risk transformation doesn't lose information — it discards the information that was confusing the model. Simplification as amplification. The binary question “is this variable in danger?” turns out to contain most of what the continuous variable knows about recessions.