# Pre-registration: Supertrend(10, 3) against a moving average, on crypto

**Registered:** 2026-11-05, before any run of this comparison and before any moving-average rule had been computed on the study's data. Not edited afterwards; any deviation is logged in a dated deviations file (`deviations-supertrend-vs-ma.md`), never silently changed. The commit that adds this file and the comparison's code is pushed to the public repository before the run, and the run's output records that commit.

## Why

"Supertrend or a moving average: which is better?" is a question people ask. The site's guide to trend reversals compares the families of method by how they work and says it has measured only one of them. This comparison measures a second on the same data, so that the page which answers the question has numbers under it, whichever way they fall.

## What is known at registration

- **Known:** every result of the Rule v1 study (`preregistration-crypto-study-2026-10-12.md`, published at `/study/rule-v1-crypto`), including its Core basket, period, trade count and drawdown.
- **Not known:** anything about a moving-average rule on this data. No such rule has been run on the snapshot, in the harness or outside it. The comparison's code is tested on synthetic series only until this file is committed and pushed.

## The rules compared

All are long or cash, read on the completed daily UTC candle, with no leverage and no shorting.

- **R, Rule v1:** exactly as live and as in the study. Supertrend(10, 3), source (high + low) / 2, Wilder ATR with a simple-average seed.
- **M(N), price above its simple moving average:** SMA_N(t) is the arithmetic mean of the N daily closes ending at candle t. Long when close(t) > SMA_N(t), otherwise cash. Equality is cash. No position until N closes exist.
- **X, the 50/200 crossover:** long when SMA_50(t) > SMA_200(t), otherwise cash. No position until 200 closes exist.

Lengths are fixed here and nowhere else:

- **Conventional set, reported in full:** N = 20, 50, 100, 200, and X.
- **Matching grid:** N = 10, 20, 30, … 250 (25 lengths).

The 10 and 3 of Rule v1 were not tuned on this data (see the study's pre-registration). The moving-average lengths are the conventional ones and a regular grid; none is chosen for its result.

## The headline comparator: the turnover-matched moving average

A fast rule and a slow rule differ mostly in how often they trade. To compare where the line is drawn rather than how often it is crossed, the headline comparator **M\*** is the grid length whose number of Core round trips over the study period is closest to Rule v1's, counted exactly as the study counts them (a round trip counts when the asset was a Core member at entry and the entry falls in the period; base costs). A tie goes to the longer length. **The selection uses trade counts only**; no return, drawdown or other outcome enters it.

If M\* falls on the edge of the grid (N = 10 or N = 250), it is reported as found and the result carries the caveat "matched at the edge of the grid".

## Everything else: identical to the study

The same frozen snapshot (SHA-256 `fd1d67415b8f8408c6fe999f5cd1fd49a1950f8ec90810205f144b370ebe39c3`), the same point-in-time Core basket and member list, the same period, the same simulation code: signal at the close of candle t, execution at the open of t+1; cash earns zero; costs per side 0.30 % base and 0.60 % stress; per-asset sleeves averaged daily to equal weight. The same metrics code.

Core members have at least 365 days of history before they join, so every rule here is past its warm-up before its first Core day.

## Acceptance test (blocking)

No results are produced unless both hold:

1. The snapshot's hash equals the one above.
2. The harness, on that snapshot, reproduces the published study's Rule v1 Core headline exactly: period start, end and number of days; compound annual return; maximum drawdown; Ulcer index; number of round trips.

## Questions

**Primary.** Q1: on the Core basket at base costs, was Rule v1's maximum drawdown smaller than M\*'s?

**Secondary, all reported.**

- Q2: compound annual return of each.
- Q3: experience of each: share of round trips that lose, average win and loss, share lost within ten days, longest run of losing round trips, average days held, time in the market.
- Q4: for each Bitcoin drawdown of at least 30 % (the study's episodes), the share of the basket's fall avoided and of the rebound missed, by each rule.
- Q5: agreement: the share of Core asset-days on which R and M\* hold the same position.
- Q6: the conventional set (N = 20, 50, 100, 200, and X) on the same measures, with each one's trade count. A disclosure table, not a search.

## Robustness

1. Stress costs, and execution one day later, for R and M\*.
2. Time halves, split at the midpoint of the period as in the study.
3. Date-block bootstrap of the paired daily returns of R and M\*: blocks of 10 to 20 days, 2,000 resamples, seed 2026 (generator seeded as in the study). 95 % intervals for the differences R minus M\* in maximum drawdown, compound annual return and Sharpe ratio.
4. Neighbours of M\*: the grid lengths 10 days shorter and 10 days longer are reported beside it, so that M\* can be seen not to be an isolated result.

One primary hypothesis, so no multiple-testing adjustment; the conventional table and the neighbours are descriptive.

## Claim ladder (wording fixed before results)

Let D be M\*'s maximum drawdown minus Rule v1's, both as positive depths, in percentage points; D > 0 means Rule v1 fell less. The difference is **clear** when all four hold:

1. |D| is at least 5 percentage points;
2. the bootstrap's 95 % interval for the drawdown difference excludes zero;
3. the difference has the same sign in both time halves;
4. the difference keeps its sign at stress costs.

| Result | The sentence |
|---|---|
| **S.** Clear, and Rule v1 fell less | "In past data Supertrend(10, 3) had smaller falls than a moving average that traded as often." |
| **M.** Clear, and the moving average fell less | "In past data a moving average that traded as often had smaller falls than Supertrend(10, 3)." |
| **N.** Not clear | "In past data Supertrend(10, 3) and a moving average that traded as often were not clearly different in how far they fell." |

**The return clause**, appended in every case, is chosen by rule: if the bootstrap's 95 % interval for the difference in compound annual return excludes zero, "Supertrend's return was higher." or "The moving average's return was higher."; otherwise "Their returns were not clearly different."

**Not allowed in any case:** "better", "best", "beats"; any forecast; any statement about moving averages or Supertrend in general, about other lengths, other markets or other periods. The numbers of both rules are shown beside the sentence. **Published either way, on the same page and with the same prominence.**

## Limitations to state

Everything the study states: crypto history is short with few full cycles; the assets move together, so the effective sample is far smaller than the asset count; assets that delisted are not in the data; results describe the past, and real execution, fees and discipline differ from a simulation. And for this comparison: one kind of moving average (simple, price above it) at lengths fixed in advance, and one crossover; exponential and other averages are not tested; matching by trade count is one definition of a fair comparison, not the only one; the two rules hold the same position on most days (Q5), so any difference comes from a minority of days; one market and one period; nothing here says which rule will fall less in future.
