Reading 44 years of Stocks & Commodities

Technical Analysis of Stocks & Commodities has been published monthly since 1982. I’ve been working with a machine-readable version of the whole run: roughly 4,000 transcribed articles by 1,360 authors, with the equations, tables, and code listings kept intact rather than flattened into prose.

A magazine is written to be read one issue at a time, and mostly to be thrown away afterwards. Reading 44 years of it in one go turns it into a different kind of document. It becomes a record of how a community of practitioners thinks, what it forgets, what it keeps rediscovering, and what it has never once got round to doing.

This is the same sort of exercise as the academic paper review , pointed at a much older and much less tidy corpus. Seven things stood out.

The technology cycles but the questions don’t

Two curves make the point. Neural networks first appear in the magazine in 1988 and take off through the 1990s, with 43 articles in that decade.¹ Then they go almost silent through the AI winter, and come back under the name “machine learning” with 29 articles so far in the 2020s. Genetic algorithms are a cleaner specimen: 21 articles in the 1990s, and after that nothing worth counting.

So the readership has been through two complete cycles of enthusiasm for computational intelligence, and both arrived with the same promises attached.

The questions being asked in 2026, though, are the questions of 1984. When is a market trending and when is it chopping? Is this move momentum or noise? The tools get replaced about once a decade and the problems never do. Anyone who mistakes a new tool for a new problem is signing up for the third cycle. The authors whose work has lasted did something less exciting: they pointed each new tool at the old problems and reported what happened.

Underneath most of it is one trade-off

If you strip the labels off four decades of indicators, then most of them turn out to be working on the same thing, which is lag against noise. More smoothing means you see the trend later; less smoothing means you trade the noise. The moving-average tutorials of the eighties, the double-smoothing constructions of the nineties, the digital-filter era of the 2000s, and a 2026 adaptive smoother that shortens its own period when the market speeds up are all one research programme wearing different clothes.

The most prolific technical author in the archive, John Ehlers, spent 120 articles over 41 years on exactly this. From 1985 onwards it’s one continuous programme: cycle measurement, Hilbert transforms, the SuperSmoother family, and this September an Adaptive SuperSmoother. His last word on the subject is a variation on his first, and I don’t mean that as a criticism. It’s a hard problem and he kept at it.

The practical consequence is a rule of thumb. Any new smoothing or trend-detection method someone hands you is, with high probability, a re-derivation of something in this pile, and it’s worth checking before you get attached to it.

A few dozen thinkers and a thousand donors

Of 1,360 named authors, 829 published exactly once. That’s 61%.

Continuity comes from a few dozen serial contributors. John Ehlers has contributed 120 articles on signal processing and related techniques. Perry Kaufman wrote 62, covering adaptive moving averages, slope divergence, high-momentum regimes, and gap momentum, always with a test attached. Thomas Bulkowski authored 53, measuring chart patterns one at a time: double bottoms, head-and-shoulders, bump-and-run, and the win and loss statistics for each. Add a handful of others and that’s the core. Their work has been mined to exhaustion. Everyone has read it, everyone has coded it, and every parameter in it has been fitted and refitted.

The one-shot majority is a different matter. Each of those 829 people left behind a single system, tested in one market over one stretch of years, and then nobody ever went back to it. Nobody replicated it. Nobody refuted it either. Taken together, I think the long tail may be the largest corpus of untested trading hypotheses anywhere, and if you’re looking for ideas that aren’t crowded, the bylines you’ve never heard could be the place to look.

Print lags the market by years, and does so reliably

New market instruments reach the archive well after they reach the world. ETFs are first covered in 2000 and don’t dominate until the 2010s, when 174 articles mentioned them. Bitcoin launched in 2009 and gets its first mention in 2014. Candlestick charting, a technique that is centuries old, appears in 1986 and takes until the 2010s (269 articles) to saturate the conversation.

Two things follow. Whatever practitioners are publishing heavily right now is already mid-adoption, so the crowding has started. And whatever is missing from the current issues is where the next decade’s literature might get written.

It’s a proposal culture

Across four decades the archive holds thousands of performance claims and very few replications. Critiques are so rare that they’re memorable. There’s one statistical takedown of divergence analysis that sticks in my mind mainly because nothing else like it exists. The 1988 system was never re-tested in 1998. Last decade’s parameters were never audited against this decade’s data. Nobody’s job was to do it, so nobody did.

That gap is the opportunity. Every article states its rules and reports its results, and every article carries a publication date. That makes the whole corpus one enormous out-of-sample experiment that has never been run. The method would be as follows: implement the rules exactly as printed, treat the author’s own period as in-sample, and let the decades after publication return a verdict. The literature supplied the hypotheses 40 years ago but, as far as I know, a large-scale evaluation that tested the real value of these thousands of articles has not been done (or least not publicly disclosed).

The discipline articles don’t need updating

The psychology and money-management pieces from 1985 and 2025 are interchangeable. Cut losses, size positions, distrust your own excitement, follow the plan. In 44 years the markets went from floor trading to microseconds and from mutual funds to perpetual futures, and the advice to the trader remains the same.

I draw a fairly blunt conclusion from that. Whatever edge lives in behaviour and discipline is permanent, because the thing it depends on doesn’t move. Whatever edge lives in information decays at about the rate the publication lag above suggests, which is to say within a few years.

Watch the code as well as the ideas

The listings tell their own history. It was BASIC in the eighties, spreadsheet formulas in the nineties, EasyLanguage for two decades after that (and thankfully now effectively dead), and by 2026 the magazine is printing Python code. Each transition lowered the barrier between reading an idea and testing it, and each one moved the practitioner’s edge somewhere else: first to being able to compute by machine, then to accessing better-quality data, then to owning a platform, and now, with computation close to free, being able to validate ideas at scale.

The 580 articles that contain author-written code are also the archive’s most honest layer. Prose can exaggerate, but the code listing either does what the author says or it doesn’t, and is the source of truth.

Why any of this is visible

The thing that made these patterns observable deserves saying on its own. That is, the literature is data. One article gives you an idea. With 4,000 of them, dates, authors, formulas, and claims intact, you get things you can actually measure: rediscovery rates, diffusion curves, hype cycles, alpha decay, and the size of the verification gap.

The practitioners of the last 44 years did the expensive part. They wrote down thousands of testable statements about how markets behave and put their names to them. What they didn’t have is what we now have: the means to hold all of it in view at once. All that remains now is to see what we can make of it all.