The 200 EMA cloud tells you where you are. It doesn’t tell you what’s next.

The 200-period moving average is the most-watched line in trading, and serious people have used it for decades. This drop does not dispute that. It tests one narrow use of one specific construction, and reports what happened.

The object

The cloud is two 200-period exponential averages on the daily chart: the upper tracks highs, the lower tracks closes. Price is then in one of three states — above the band, below it, or inside it. The band is asymmetric by construction and narrow: median width 0.57% of price.

Two mechanical claims get made about it: that the state tells you which way to lean, and that the flip is worth acting on. We pre-registered both, then tested them.

The denominator nobody quotes

Every session is an event. Enter at the close, put a barrier ±0.5×ATR10 either side, ask which is touched first within ten sessions. Symmetric barriers on a near-random walk should pay a coin flip, and they do: 50.21% favourable, n=3,836, mean R +0.004, zero timeouts.

50.2% is the number to beat. Not 50, not 60.

What it looks like it says

An eight-point spread. It is the fifteen-year uptrend wearing an indicator’s clothes. Price sat above the cloud on 82% of sessions in this sample and NQ rose roughly tenfold across it, so shorting anything in those years loses. The correct comparison is not the short leg against the long leg — it is each leg against the drift it inherited. Against that, the spread goes away.

The walk-forward

Four purged expanding folds with a 200-session embargo, equal to the 200-day lookback, so no test fold shares average-history with the data that trained it. Registered in advance: beat the benchmark by a Brier improvement of 0.002 in at least three of four folds.

One fold of four cleared the bar; the rule needed three. Overall the conditioned signal came in at 49.4% against the 50.2% benchmark, mean R −0.013.

The placebos are the actual finding

We reran the identical test on three deliberately broken copies: one frozen five sessions in the past, one with its sign inverted, one cheating with tomorrow’s state.

Read the third row again. Running the signal backwards cleared the threshold more often than running it forwards. A week-old copy improved more folds than the live one. Even the arm handed tomorrow’s answer could not use it.

Inverting the signal cleared the bar more often than the signal did.

This is the part worth keeping, and it is a statement about timing information at this horizon: a variable you can freeze for a week or run in reverse without changing the outcome is not telling you when. It is telling you where. Those are different jobs and the cloud only does one of them.

The null agrees

We permuted the state in 200-session blocks — blocks, not days, because the state persists for months and a naive shuffle would fake a tight null — and reran the whole walk-forward 200 times. Observed improvement −0.00033; the null’s average improvement −0.00042. p = 0.56. A reshuffled cloud beat the real one on average.

98 opinions in fifteen years

The flip claim could never have been settled here, and understanding why applies to every slow indicator you will be shown.

In 3,847 sessions the cloud flipped 98 times — 46 up, 52 down — and the state resolves into just 155 contiguous regimes. Days inside one regime are not independent observations: a 300-day stretch above the band is one observation, not 300.

We wrote the power limit into the record before running anything. With 46 up-flips the smallest detectable effect is a jump to roughly 65%. Nothing built on a 200-day average is right 65% of the time, so this sample could not have found a real effect of plausible size. We declared that first so a null could not later be sold as a discovery, and a lucky hit could not be sold as proof. For the record the flips came in at 45.9%, mean R −0.082.

What this does not test

This matters more than the result, so it gets its own section rather than a footnote.

• Longer horizons. Ten sessions. The best-documented use of the 200-day average operates on months, and a variable with no ten-session timing content can still carry multi-month information.
• Exposure management rather than direction. The canonical published result — Faber’s 2007 tactical-allocation work, tested across asset classes back to 1972 — is a rule that holds the asset above its 10-month average and moves to cash below it, checked monthly. Its reported benefit is equity-like returns with bond-like volatility and drawdown. That is a risk claim, not a direction claim, and our test measured the wrong thing to speak to it. Note especially: that rule goes to cash below the average. We went short — a trade it would never take, in the sample where it loses most.
• Other instruments. NQ only, and a fifteen-year sample that is overwhelmingly one regime.
• Discretionary use. A trader using the cloud as one input among several — for context, for sizing, for deciding what not to do — is doing something this test cannot evaluate. Mechanising a discretionary read and killing the mechanical version does not kill the read.
• Simple vs exponential, and other constructions. We tested this cloud. Most of the literature uses a simple average on monthly closes.

So: it is entirely possible to trade profitably with the 200-period average. Nothing here argues otherwise, and we would not have the evidence to. What failed is the specific mechanical thing — symmetric long/short direction from the cloud state, on NQ dailies, over ten sessions.

What we keep

The cloud is a good description. It says where price is relative to a long average, it says it unambiguously, and a room full of traders reading it will agree on what they see — which has real value. What it did not do, in the one use we measured, is tell us what happens next.

Where you are is not where you are going. Most of what a chart shows is a summary of the past presented in the grammar of a prediction. Knowing which one you are holding is the whole job.