Ichimoku means one look.
Built in 1930s Tokyo by a newspaper writer under the pen name Ichimoku Sanjin — “a glance of a mountain man” — its promise is right there in the branding: five lines, one shaded cloud, and the market’s whole condition legible in a single glance. It is, I think, the most beautiful thing in retail technical analysis. It is also the perfect specimen for this series, because beauty and knowledge are exactly the two things a simulacrum invites you to confuse.
So I took one honest look — sixteen years of daily Bitcoin, 2014–2019, with 2020-onward sealed. Three findings fell out, and each one deflates the last.
One — The machine is simpler than it looks#
Strip the mystique and Ichimoku is a moving-average system in a costume — its own primary source calls it “similar to moving average techniques” and plots the control to prove it. Worse, its sacred constants are a payroll calendar: the famous 26 is the number of trading days in a month “Saturdays included,” from when Tokyo traded six days a week. And one of the five celebrated lines, Senkou Span A, is algebraically just the average of two lines you already have — I measured its independent information at exactly 0.0000000000. The five-line “confluence” everyone waits for isn’t five witnesses. It’s double-counting.
Two — The edge was six trades#
Backtested honestly, cloud-following does beat buy-and-hold — until you look at where the money came from:
Six trades out of thirty-five produced +623.9%. The other twenty-nine collectively lost 17.6%. Remove the top six and the whole edge goes negative.
You’re not betting on Ichimoku. You’re betting that Bitcoin keeps producing 50–200% parabolas — a far shakier assumption in a post-ETF market than in 2016. The Sharpe “improvement” (0.87 → 1.23) is comfortably inside its own ±0.5 error bars. And the same rules that make 19x on a low-fee perp lose to buy-and-hold on retail spot. The venue matters more than the signal.
Three — You can’t filter a fat tail#
The reflex every quant has within ninety seconds: use machine learning to keep the six good trades and skip the junk. It cannot work here, and the reason is arithmetic, not data. A flawless filter of the losers gains you +17.6% over six years; missing one tail winner costs −213.5%. So the moment your classifier is anything short of perfect, the math inverts:
At 95% recall — an outstanding classifier — the filter is negative in expectation while looking like it works 74.4% of the time. A machine for fooling yourself. Negative skew doesn’t announce itself; it waits.
The general lesson, which reaches far past Ichimoku: for fat-tailed payoffs, precision-improving machine learning destroys value. It optimizes the thing that doesn’t matter (small losses) at the cost of the only thing that does (never missing the tail). The fix isn’t a better filter — it’s sizing: never go binary, scale in as the trend survives.
What survives the autopsy is real but small: the cloud as a regime gate, kumo thickness as a volatility read, the Kijun line as a mechanical trailing stop. What doesn’t: the forward cloud as a forecast, “kumo twist” turning-points, Chikou “confirmation,” and 9/26/52 as anything sacred.
The cloud is beautiful. It renders gorgeously. It makes a chart look like a system, and a system look like knowledge. That gap — between looking like knowledge and being knowledge — is where the money goes to die. Ichimoku means one look. It turns out one look is exactly the problem.
📄 The full analysis is a downloadable paper — every table, the derivations, the degrees-of-freedom ledger, the Monte Carlo, the trade-level teardown, and the single pre-registered test I’d stake it on:
→ One Look at Ichimoku — Full Analysis (PDF, 10pp)#
None of this is advice — it’s a methodology exercise on a public price series.
Next in this series: why your backtest lies — sample size, multiple comparisons, and the arithmetic of getting fooled.