RSI — The Parameter Dossier

R.1The defendant

Three parameters, and each has a different pedigree. The period and zones started life as pure folklore; Round 4 re-tuned them under exam conditions. The exit bracket (ATR×1.5, 1:2) is the roster-standard risk suit, never tuned per-strategy. This dossier re-litigates all of it with a 70/30 split: tune only on data before 2024-03-26, judge only on data after.

R.2Where the defaults came from — and how the classics failed

RSI's famous numbers — period 14, zones 30/70 — come from Welles Wilder's 1978 book. They are not laws of nature; they are what one engineer liked on 1970s commodity charts. Here is what they do on GBPUSD H4, against the Round-4 tune:

So yes — the original defaults were wrong, at least for this pair on this timeframe. Not scandalously wrong; wrong the way a borrowed suit fits: it was measured for someone else. Round 4's re-tailoring is what earned RSI its roster seat.

R.3Does GBPUSD's character explain it? (Honest answer: no)

The folk theory says a rubber-band strategy wants a mean-reverting market — one where an up-move makes a down-move slightly more likely. We can measure that: the lag-1 autocorrelation of H4 returns (negative = snap-back tendency), alongside drift and choppiness:

Read that twice, because it demolishes the folk theory: the most snap-backy pairs are where RSI loses, and the least snap-backy, hardest-drifting pair — USDJPY — is where it does best. The character numbers are also tiny (±0.02 on a scale where 0 means coin-flip): at H4, no major is meaningfully mean-reverting. The honest conclusion, same as Sentinel's Part C: a pair's coarse "character" does not predict where an edge lives — the exam decides, not the theory. That is exactly why we test per pair instead of assigning strategies by folklore.

R.4The grid trial — plateaus, and a beautiful mirage

Sixteen configs (period × zones) fought on the tuning half only. The in-sample beauty contest:

Period 26 with zones 35/65 glows brightest — PF 1.38, nearly +30%! Surely better than our 1.16? Now the exam half:

Verdict on the roster numbers: keep period 20, zones 35/65. They sit on a modest plateau (the 20/35–40 corner holds 1.15–1.16 IS), and they alone survive out of sample. No change.

R.5The formula affair — our RSI is not MT5's RSI

While reading the code for this dossier, a discrepancy: the docstring says Wilder's smoothing, but the code computes a plain rolling average — that's Cutler's RSI, a legitimate variant, but a different number from what MT5 draws. How different? Averaging 5.9 RSI points apart, sometimes 14+; they agree on "price is in a zone" only 39% of the time.

Both versions make money — the idea is robust to the formula, which is genuinely reassuring. But Wilder's is choosier: half the trades, two-thirds winners, drawdown nearly halved, and slightly better out of sample. Two consequences. First, a practical warning: when you check RSI on an MT5 chart, expect its value to differ from ours — that is arithmetic, not a bug in either. Second, a candidate improvement: switching the engine to Wilder's smoothing would match the platform and looks better risk-adjusted — but it changes live behaviour, so it goes on the docket for a proper sealed exam, not a quiet code edit.

R.6The USDJPY surprise

The transfer test — GBPUSD's tuned config, applied unchanged to every major — mostly showed red, as expected. Except one row:

USDJPY beats RSI's own home pair — in both halves, with 258 exam trades and the smallest drawdown on the board. And note the delicious irony: these zones were tuned on GBPUSD, so for USDJPY this is doubly out-of-sample — no part of this config ever saw a yen candle. It echoes the edge map's lesson that USDJPY is simply fertile ground, and it makes rsi · USDJPY H4 a genuine candidacy. Candidacy, not coronation: before it touches the roster it must pass its own Round-4-style sealed exam, the same gate every roster member walked through.

R.7Dossier verdict

Next dossier: RevertX — six pairs, zero tuned parameters, and a baseline resting on only 14 trades.