Four alerts in one gold move. How many tests of a signal?

First decide what one observation means

If I ask what a subscriber received, each distinct alert belongs in the record with its own timestamp, levels and model version. If I ask how often an idea worked in a different market episode, the shared period of falling gold matters. Several alerts may depend on the same minutes of price history for their classification. They can be valid separate messages while carrying far less evidence about repeatability than the same number of alerts across unrelated sessions.

Cameron and Miller describe the statistical problem of observations that move together within groups. Their work concerns inference from grouped data. Applying the idea to alert journals requires an explicit choice about which records might share a shock. A different model name or a different entry minute cannot establish independence.

Source notes: [1] A. Colin Cameron and Douglas L. Miller — A Practitioner’s Guide to Cluster-Robust Inference

One busy episode can decide the headline

Here is a hypothetical example, with no CY66 records or market prices. In episode A, four eligible alerts each meet a predefined test. In episode B one alert misses it, and in episode C one alert misses it. Counting every alert gives four successes in six observations, or about 67%. Giving each episode equal weight means averaging the three within-episode success shares of 100%, 0% and 0%. That produces about 33%.

Both calculations use the same rows. The first describes the share of alerts that met the test in this invented file. The second describes the mean success share in a typical episode when each episode gets one vote. Choosing between them changes the question. Neither estimates the future success rate of a gold strategy from three episodes.

Define the overlap before looking at outcomes

For a study that measures the next hour after every alert, write down each one-hour evaluation window in UTC. One reproducible grouping rule joins alerts on the same instrument whenever their windows overlap, including chains in which one alert overlaps two neighbours. Save the first and last time and the member IDs for every resulting group. A chain can stretch across a long session, so publish the distribution of episode lengths and explain any maximum duration chosen in advance.

Run the calculation twice, once per alert and once per episode. Then remove the busiest episode and recompute both descriptive results. In the hypothetical example, removing A leaves two misses. That sensitivity tells the reader exactly where the favourable observation came from. It also identifies a concrete research priority: collect other episodes under the original rule, including quiet periods when few alerts arrive.

Keep an honest measure of precision

Correlated observations affect how precisely an average can be estimated. NIST’s work on repeated measurements demonstrates the issue for time series. In a signal study, treating each overlapping row as an independent draw can make an interval look more precise than the evidence warrants. Moving to episode counts also requires judgment because neighbouring episodes can share a longer price trend.

I would publish the raw count, the episode count, the concentration in the largest episode and the grouping rule before attempting an interval. With three episodes I would stop at those descriptive results. A reader can see what happened within the sample and what new observations would make the test more persuasive.

Source notes: [1] A. Colin Cameron and Douglas L. Miller — A Practitioner’s Guide to Cluster-Robust Inference · [2] Nien-Fan Zhang, NIST — Calculation of the uncertainty of the mean of autocorrelated measurements