Signal Decay Visualizer

A measure of how well a signal's predictions line up with subsequent returns, calculated as the correlation between the signal's values across a universe and the realized returns over the following period. It runs from minus one to one, and in practice useful equity signals produce small positive values. Its size matters less than its consistency, since a small edge applied repeatedly across many positions is what a systematic strategy relies on.

How quickly the signal's predictive relationship with future returns fades as the horizon lengthens. A short half-life means the information is priced in quickly, so acting on it requires trading soon after the signal appears and repeating that often. A long half-life means the relationship persists, which allows lower turnover. The figure is therefore a direct input to how often a strategy has to trade, which in turn determines how much cost it must overcome.

Because a raw coefficient series is noisy enough that any two adjacent points can point in opposite directions. Fitting a curve summarizes the whole series into a rate and a starting level, which is a comparable pair of numbers rather than a shape to be interpreted by eye. The cost is that a fit imposes a functional form: a signal whose decay is not exponential will still produce a half-life, and that number will describe the fit rather than the signal.

That the signal ranked in the opposite direction from subsequent returns over that period. A single negative reading is usually noise. A persistently negative series means the relationship runs the other way, which is a finding rather than a failure, though it needs the same validation any positive result would get before being used inverted. The common cause of a sign flip is a definitional error in how the signal or the return was constructed.

Enough that the fit is not dominated by a handful of observations, and the requirement grows as the coefficient values shrink toward zero. Because typical values are small relative to their variation, a short series produces a half-life with very wide uncertainty around it. Treating the fitted number as a point estimate rather than as the middle of a range is where the estimate is most often overinterpreted.

To separate the trend from period-to-period noise. The raw series carries the sampling variation of each individual period, which can obscure a gradual decline. A rolling mean removes much of that at the cost of lagging genuine changes, so a shift in the signal's behavior appears later in the smoothed line than in the raw one. Reading both together shows the direction and how confidently it can be asserted.

It sets the upper bound on how long a position can be held before the reason for holding it has faded. Rebalancing much less often than the half-life means holding positions whose signal has largely expired. Rebalancing much more often than the signal changes generates turnover and cost without new information. The practical rebalancing frequency sits between those, adjusted downward for trading costs, which push toward holding longer than the signal alone would suggest.

Adoption is the usual explanation: as more participants act on the same relationship, it is priced in faster and the half-life shortens. Market structure changes have a similar effect, since faster dissemination of the underlying data compresses the window. Regime changes can shift it in either direction. Re-estimating the decay periodically, rather than fixing it once, is what surfaces a change before the strategy's rebalancing frequency has become wrong for it.

It makes a cheaper one to run, which is not the same thing. A slow decay allows lower turnover and therefore lower cost, but it says nothing about the size of the edge, its consistency across market conditions, or how many positions it can be applied to. A fast-decaying signal with a large coefficient can outperform a slow one with a small coefficient, if the strategy can trade quickly enough and cheaply enough to capture it.