Matching Data Frequency to Signal Frequency

A quarterly earnings-revision dataset updates four times a year. A daily trading signal needs a value every day. Somewhere in between, someone has to decide what the signal does on the 88 days out of 91 when nothing new has actually arrived — and that decision, more than any modeling choice, determines whether the signal is measuring real information or just repeating old news with a fresh timestamp.

The stale-fill step function

The common approach is a stale-fill, or last-observation-carried-forward: between two quarterly updates, the daily value used is simply whatever the most recent quarterly reading was.

$$x_t = x_{\tau(t)}, \quad \tau(t) = \text{last update date} \le t$$

In words: today's value of the signal equals whatever the underlying data said on the most recent day it was actually updated, no matter how long ago that was. This is honest as long as the rest of the research pipeline knows the signal is a staircase and not a smooth series.

Worked example

A researcher builds a daily cross-sectional signal from quarterly consensus-EPS revisions and finds a Sharpe ratio of 1.4 in a naive daily backtest, rebalancing every day. When the same signal is instead rebalanced only when the underlying data actually updates — roughly every 60 trading days per name, staggered across the universe — turnover falls by a factor of 15 and the Sharpe ratio, net of realistic costs, drops to 0.6. The apparent daily edge was mostly the backtest rewarding itself for holding a static position and calling every unchanged day a fresh, independent bet.

What this means in practice

Match the rebalance frequency of a signal to the true update frequency of its slowest input, or explicitly blend a slow input with a genuinely fast one (like price) so the combined signal has a real reason to move daily. Going the other direction — down-sampling a fast feed to a slower cadence, say averaging tick data into weekly bars — throws away information but at least does no active harm.