Why 30 daily Bitcoin forecasts are not 30 independent tests
A timeline example showing overlapping horizons, matched baselines, and why the size of a forecast ledger can overstate the strength of evidence.
Published by Marcos Mendes · BTC Horizon
Count forecasts, then inspect their dates
A daily 30-day forecast issued on October 1 targets October 31. The edition issued on October 2 targets November 1. Their observation windows share 29 days. Publishing another edition adds a row, but it does not create a wholly new market experience.
This is a hypothetical timeline, not a claim about October performance. It illustrates why a ledger with 30 resolved forecasts can contain substantial dependence: a single broad market move can influence many of the outcomes at once.
What a non-overlapping comparison changes
One simple diagnostic is to select a forecast, then select the next one only after its target timestamp. With a 30-day horizon and daily issuance, this yields roughly one observation per month. Choose the selection rule before inspecting performance, and keep it the same for the candidate and its baseline.
Non-overlapping windows reduce one obvious source of dependence; they do not guarantee independence. Volatility regimes and market trends can persist across adjacent months. The smaller count is a reason to express uncertainty rather than to switch back to a more flattering sample.
Match the baseline to the same outcomes
For a yes-or-no direction event, a constant probability of 0.5 scores 0.25 under the Brier score on every observation. A model averaging 0.24 has a lower point estimate, but a small or dependent sample does not by itself establish a dependable advantage.
Compare forecasts on the same target dates, outcome source and validity rules. Dropping difficult dates only from the candidate’s sample or comparing its recent period with a baseline from another regime answers a different question.
A practical reading order for BTC Horizon
First identify the model and source in the methodology. Then check how many forecasts have matured, how many were invalid, and the dates of the available sample. Read direction scores alongside interval coverage and price error. Where overlapping and non-overlapping results are available, report their counts separately.
An absent comparison is a limitation to disclose, not a result to infer. This article explains an evaluation design; it does not claim the current CMC edition has passed it. A longer ledger becomes useful evidence only when its publication and outcome rules remain inspectable.
Methods and evidence
These notes explain BTC Horizon’s implementation and evaluation rules. Worked examples are illustrative unless explicitly described as observed records. Articles are prepared with AI assistance using the project’s code and documented records; they are not independent peer review.