A Match Is Not a Forecast

Two lines sit on a screen. One is dark, the other bright. They begin apart, cross, separate, then curl back toward each other. After a little stretching of the axes, the resemblance becomes hard to miss.

The eye does the rest.

Once the lines have met often enough, the older one starts to feel like a guide. Its unfinished twin seems to be following a route already drawn. You know, in the abstract, that a chart cannot remember. Yet the continuation on the old line now looks less like history and more like an appointment.

This is why market analogy charts work so well as images. They turn uncertainty into a shape.

They have always irritated me. I have regarded them as useless, distracting and unserious noise, especially when a dramatic old crash is left hanging beyond the point where the current line stops. That may be too absolute. Past prices can contain information. The serious question is whether a selected historical twin contains any.

What would have to be true for that old line to help?

A recent post offered a neat example. After 170 trading days, it said, the S&P 500 in 2026 was tracking 1978 more closely than any other year. Both were midterm years, both were described as facing another inflation wave, and both were years of the horse. The heading was “Only for bears.”

It is an excellent piece of market storytelling. The price match arrives first. The macro details then make the match feel causal. The old chart supplies the ending.

But the method has two missing denominators.

The first is the search. How many starting dates, historical years, indices, scaling choices and similarity measures were available before the winner appeared? Correlation of price levels, correlation of returns, volatility adjustment and a stretched time axis can all choose different twins. Even a fixed rule will always produce a closest year. “Closest” does not tell us whether the distance is unusual or whether the winner beat the rest by anything meaningful.

Without the full search record, the reported correlation has no calibrated meaning. The same problem sits underneath the small-sample FinTwit statistics I wrote about recently, but an analogy chart makes it harder to see because the selection looks like geometry rather than sampling.

The second missing denominator is worse. The chosen analogy gives us one sequel.

One path after one historical match is not a forecast distribution. It has no base rate, no range and no honest error bar. If the old market fell next, the chart looks bearish. We are rarely shown all the nearly as good matches that rose, drifted sideways or broke down at a different point. The picture replaces that absent sample with visual confidence.

There is a respectable version of this idea. Define the distance measure before looking at outcomes. Search only a training sample. Take many nearest neighbors rather than the prettiest one. Compare all their subsequent returns with an unconditional benchmark, then freeze the method and test it on new data. If the result survives costs and reasonable changes in parameters, you may have found a weak signal.

At that point, however, you no longer have an analogy chart. You have a model.

Finance has tested versions of that model. Francis Diebold and James Nason used locally weighted, nearest-neighbor-like methods on ten major dollar exchange rates from 1973 to 1987. The nonlinear forecasts did not improve out-of-sample point prediction over the simple benchmark. That is close to the analogy claim in operational form: find similar past return states, weight their continuations and see whether the forecast improves.

The counterweight matters. Andrew Lo, Harry Mamaysky and Jiang Wang later defined familiar chart patterns with an algorithm and applied them across hundreds of stocks. Some patterns changed the distribution of later returns, especially among Nasdaq stocks. They concluded that several patterns carried incremental information. They did not conclude that this guaranteed excess trading profits.

So price history is not blank. Momentum, volatility clustering and some tightly defined patterns may carry information. A visual match can accidentally proxy for one of them.

But if the claimed reason is inflation, policy, positioning or forced selling, price resemblance alone has not identified it. Those are extra state variables. Adding them after selecting the matching year creates a persuasive story, not a price-only test.

This is where I would soften my old view and sharpen it at the same time. Analogy searching can be useful as hypothesis generation. It can suggest a regime variable or a pattern worth defining. As published evidence for a directional trade, the familiar single-twin chart is close to worthless.

I also think it can harm novice investors. Crash analogies borrow authority from a disaster everyone remembers while hiding the many false alarms nobody archives. They make fear look measured. A reader receives a line, a date and a historical rhyme, but none of the information needed to judge the forecast.

The next market observation will arrive after the trade. The analogy offers no second sequel to keep the first one honest.