Ichimoku Cloud Strategy

From a chart idea to a rule a bot can run

The basic definitions of the five lines are on the Ichimoku cloud page. Here is the short version, only so the rules below make sense. The tenkan is the midpoint of the highest high and lowest low of the last 9 bars. The kijun is the same over 26 bars. Span A is the average of those two, drawn 26 bars ahead. Span B is the midpoint over 52 bars, also drawn 26 bars ahead. The chikou is the close drawn 26 bars back. The area between span A and span B is the cloud.

A person looking at a chart does not need more than that. A bot does. "Price is above the cloud" can mean the close, the high or the low. It can mean above the upper edge, or above both edges on this bar, or above the cloud that was plotted ahead of the bar. "The tenkan crossed the kijun" can mean it is above now, or that it was below on the last bar and is above on this one. Each wording gives different trades.

So the first job is to pick one meaning for each phrase and write it as a comparison on numbers. Here are the building blocks most Ichimoku rules use.

• Price versus cloud. Long filter: the close is above the larger of span A and span B on this bar. Short filter: the close is below the smaller of the two. Inside the cloud: neither, so the filter is off.
• Tenkan and kijun. Either a state (the tenkan is above the kijun) or an event (it was below on the last bar and is above now). The state is true for many bars. The event is true for one bar.
• Chikou. The close is above the close from 26 bars ago, which is the chikou test in plain numbers.
• Exit. The opposite cross, the close falling back into the cloud, or the close falling below the kijun. Without an exit rule a bot has an entry and no way out.

Two points about timing. The cloud on a bar is built from values 26 bars old, so it is known in advance and a rule using it today is not looking ahead. The chikou is different: it is plotted 26 bars back, but the value is today's close. A bot must compare today's close with the close from 26 bars ago, not read a plotted point as if it were old data. Getting this wrong is a form of look-ahead bias.

A worked example on a made-up bar

The prices here are made up. The path is 260 closes built from a slow rise plus three waves, and each high and low sits 0.8 either side of its close. All the arithmetic below was run in node on 31 Oct 2026 and is exact for this path. Take bar 120 and the standard 9, 26 and 52 settings.

Now test the two rules. Rule one: the close is above the cloud top. 102.77 against 103.31: the close is inside the cloud, so rule one is false. Rule two: the tenkan is above the kijun. 107.49 against 109.30: rule two is false. A rule that needs both is false on this bar, so a long-only version of that rule would be flat at this close.

How many rules are there, really?

Count the choices from the list above. This is a made-up counting exercise, but the arithmetic is exact.

Multiply the first five rows: 3 x 3 x 2 x 3 x 2 = 108. Remove the 6 variants with every filter switched off, because they have no entry. That leaves 102. Multiply by the five period sets and you get 510. If instead each of the three periods can be any of ten values, there are 1,000 period sets and 102,000 variants.

Why many variants invite overfitting

Suppose a bot tries all the variants on the same history and keeps the one with the best result. Some variant will always come out on top. The question is whether it came out on top because it describes something real, or because with so many tries one of them was bound to line up with the quirks of that stretch of prices. From the result alone you cannot tell. This is overfitting, and the more variants you try, the less the best result tells you.

A small honest test on the made-up path

Here is the shape of a fair test, run on the same made-up path. Split the history in two. Part A is bars 77 to 167, where all the lines exist and the settings may be chosen. Part B is bars 168 to 258, which is locked away until the choice is made. The bot holds one unit when its rule is true at a close, and is flat otherwise. Each change of position costs 0.1 points. Results are in points per unit. The settings come from four period sets and three rule types, twelve rows in all, so this is a very small search compared with the counts above.

In Part A, 5 of the 12 rows finished above zero. The best Part A row was 7/22/44, tenkan above kijun, at 8.51 points. Applied to Part B, that same row gave 20.99 points. The best Part B row was 5/15/40, tenkan above kijun, at 28.64 points. So even in a search of only twelve rows, the winner in the first part and the winner in the second part were different rows.

Look at the trade counts as well. Every row made between 3 and 7 trades in a part. With so few trades, one more or one fewer changes the total a lot, and nothing here can be called a pattern. This is a made-up path chosen to show the method. It says nothing about how any rule behaves on real prices.

How to test it honestly

• Write the rule first. Fix the wording, the settings and the exit before looking at results. Keep a list of every variant you try.
• Keep data back. Choose on one part and test once on another. The how to backtest guide shows the steps.
• Roll the window. Walk-forward testing repeats the choose-then-test step across many windows, so one lucky stretch cannot decide.
• Include costs. The cross rules change position often when the price chops sideways. Add the spread, fees and slippage for your market. They vary by broker or exchange, so check yours.
• Look beyond the total. Check the trade count, the worst losing run and the largest fall from a peak. A good total with one huge winner is fragile.
• Check neighbours. If a 9, 26 and 52 rule works and a 10, 27 and 54 rule collapses, the first result was probably a fit to noise. A rule worth studying works across a range of nearby settings.

What it does not tell you

A set of Ichimoku rules describes what the price did relative to its own past. It does not say what comes next. A backtest, even a careful one, shows what a rule would have done on data already seen. It cannot show that the same behaviour will repeat. This page is education about how to turn a chart idea into something testable, not a signal and not advice.