Curvature-penalized exponential averages versus MACD

Smoother design as a decision rule

Indicators are turned into decisions through constructions such as two-average crossovers or a threshold penetration.

Lengthening a moving average to suppress high-frequency oscillation increases lag, while shortening it to cut lag restores that oscillation. That restored oscillation is wiggle, and it makes the sign of the slope hard to read.

A curvature penalty on exponential smoothing

A standard exponential moving average updates by adding a smoothing-constant share of the gap between the latest observation and the prior average. That recursive update is exponential smoothing.

Ordinary least-squares fitting is described as capturing both signal and noise. Regularization adds a curvature penalty that serves as a control for indicator wiggle.

The regularized exponential update mixes the last two regularized values with that same observation gap, using a smoothing constant and a regularization constant, then scales by one plus the regularization constant.

Seeds, bounds, and conventional averages

The first two regularized values are typically seeded from moving averages, and how soon the series settles depends on those seeds.

Because a regularized series depends on prior curvature, it is not guaranteed to stay inside the input extrema, so conventional averages are presented as better suited to crossover and threshold rules.

Slope sign versus a MACD control

Regularized momentum is the one-step change of the regularized exponential average divided by its current value, and a zero crossing of that quantity is treated as a direction change.

Editorially, that reading is gradient logic: the hypothesis is the sign or zero crossing of the indicator's rate of change. MACD is kept beside it as a familiar moving-average convergence-divergence zero-cross control.

A visual overlay of regularized momentum and MACD showed similar smoothness, with the regularized series often crossing zero a few bars earlier.

For rules based on slope sign, regularization is presented as the better-matched construction. For crossover or threshold rules, conventional averaging is presented as superior.