The Dow Jones Industrial Average in 1896 was a collection of 12 companies trading on paper tickets in specialist pits. Communication happened via telegraph. Calculation required pencil and paper. The fastest trade execution took minutes. In 2026, the index comprises 30 companies trading in microseconds through fiber-optic networks. Algorithms execute millions of trades per second. Information travels at the speed of light. High-frequency trading firms can arbitrage price differences lasting nanoseconds.
By any logical measure, these markets should have nothing in common. The technology, participants, regulation, and information infrastructure are completely transformed. Random walk theory would predict that such radically different systems would produce radically different price patterns. The mathematics of one era should not apply to the other.
And yet: the same Fibonacci proportions that describe the 1896-1932 Inception Swing continue to describe the 2009-2026 super-cycle. The persistence is the observation this page is about. Whether it reflects something deep in market structure, or a proportional relationship loose enough to fit most price histories, is a question the evidence assembled here narrows without settling.
The interpretation we favour is that market structure operates at a level beneath the surface mechanisms we observe—that technology, news events and participant behaviour act as placeholders within a recurring geometric arrangement, while the proportions themselves change little. That is a hypothesis. It is consistent with what follows, and it is not the only thing consistent with what follows. Readers who want the strongest version of the evidence should look to the cross-market work cited in Section 3, which was conducted on mechanically detected swings across 25 international indices, rather than to the era tables below, which were assembled by hand and are small.
Let's enumerate just how different these market eras actually are:
These are not incremental differences; they are categorical transformations. The 1896 market operated at human cognitive speed. The 2026 market operates at machine speed—six orders of magnitude faster. Information that took hours to disseminate in 1896 is now globally distributed in milliseconds. The participants, the infrastructure, the regulatory framework—everything has changed.
Given this transformation, what would we expect to observe if markets follow random walk theory (the dominant academic model)?
"Price changes are unpredictable because new information arrives randomly. Since information arrival, processing speed, and participant composition have radically changed, the statistical properties of price movements should also have radically changed."
Specifically, random walk theory would predict:
This is the logical expectation. Markets are complex adaptive systems. Change the participants, change the infrastructure, change the rules—you should get different emergent behavior.
If markets are informationally efficient (prices reflect all available information) and mechanistically transformed (execution speed increased 1,000,000x), then:
This is the null hypothesis as we originally framed it, and it is framed too generously to ourselves. Random walk theory does not in fact predict that geometric ratios have zero applicability across eras: a random walk still produces swings, retracements and proportional relationships between them, and a level placed at a fixed ratio of a real swing will sometimes sit near a later turn for no better reason than that prices trend. The honest null is not "no correlation." It is "the correlation one would obtain by applying the same procedure to price series with no structure of interest," which is a considerably harder bar and the one Section 3 now uses.
Let's be even more specific. Here's what random walk theory predicts for our Fibonacci extension analysis:
If we calculate Fibonacci extensions from 1932 low (41.22) and test them against 2009-2024 price action:
An earlier version of this page set the benchmark as follows, and each figure is worth examining because each is wrong in an instructive way.
The lesson is not a small one, and it applies to any framework of this type. A model tested against a null that understates chance will appear to succeed spectacularly, and the size of the apparent success is a measure of how badly the null was specified. Exceeding a 10% baseline when the true baseline is 54% is not a discovery about markets. It is a discovery about arithmetic. Where this page previously concluded that beating the benchmark meant something other than random information processing governs price structure, the correct conclusion was that the benchmark had been drawn in the wrong place. Section 3 states what remains once it is drawn in the right one.
Now let's look at what actually happened. The following summary draws from the comprehensive forensic analysis presented in the Evidence series:
For complete historical validation and detailed calculations, see:
Note: These posts contain the raw data tables, charts, and calculations supporting the claims below. As those analyses are refined, the specific numbers here will be updated to reflect the latest findings.
| Era | Technology | Levels Tested | Hit Rate (±5%) | Median Error |
|---|---|---|---|---|
| 1896-1932 | Telegraph, specialist pits | 8 | 100% | 1.2% |
| 1932-1966 | Telephone, tickers | 5 | 80% | 3.8% |
| 1974-2000 | Computers, electronic trading | 7 | 86% | 6.7% |
| 2002-2009 | Algorithmic trading, derivatives | 3 | 100% | 3.2% |
| 2009-2024 | HFT, AI, global fiber networks | 10 | 90% | 2.6% |
| TOTAL (130 yrs) | Five distinct technological eras | 33 | 91% | 3.4% |
Across five technological regimes and 33 inflection points examined, the extension levels sat close to the turns more often than not, and the median error did not widen as the technology changed. That second observation is the interesting one, and it is the claim this page rests on: the accuracy does not decay across eras.
What these figures are, and what they are not. The inflection points in the table above were identified by us, not detected mechanically, and several eras rest on a handful of levels—three in one case, five in another. Percentages computed on samples of that size are descriptive of the cases chosen and cannot be projected onto the population of all turns. We have therefore not stated a hit rate for the corpus as a whole, nor computed a p-value against a random-walk null, and an earlier version of this page did both. The null in that calculation assumed a single level with a tolerance band; the ladder has many rungs, whose bands together cover a substantial fraction of the price range. A correctly matched baseline is far higher than the one previously quoted, and the apparent outperformance correspondingly smaller.
The proposition that survives independent testing is more modest and rests on far more data. The era tables are offered as illustration of continuity, not as its proof.
That result was produced on swings detected mechanically, scored against a baseline built to reflect how much of the price range the ladder's bands actually cover, and it was positive in every one of the twenty-five markets examined. A lift from 54% to 75% is a real and useful edge. It is not a 9x outperformance, and it never was.
The most compelling evidence comes from the most technologically advanced era—the period that should show the greatest deviation from historical patterns if technology were the dominant factor. Let's look at the specific levels:
| Level | Calculated from the 2007–09 swing | Closest approach (2009-2024) | Date | Error |
|---|---|---|---|---|
| 1.0 Extension | 14,164 | 14,054 | Feb 2013 | -0.8% |
| 1.618 Extension | 18,870 | 18,053 | Dec 2014 | -4.3% |
| 2.618 Extension | 26,490 | 26,616 | Jan 2018 | +0.5% |
| 4.236 Extension | 38,820 | 36,952 | Jan 2022 | -4.8% |
Three corrections to how this table was previously described. It was captioned as calculated from the 1932 anchor; it is not. These levels derive from the 2007–2009 swing (a decline from 14,164 to 6,547), using ratios drawn from the 1896–1932 template. The distinction matters, because extensions of a recent swing are a much less surprising thing to see respected than extensions of a swing ninety years old, and the original wording claimed the latter while showing the former.
Second, the arithmetic was fixed once the 2009 low was in place, but no contemporaneous published record of these projections exists, and the choice of which subsequent highs count as approaches was made afterwards. The honest description is a retrospective fit.
Third, the market did not "hit every level." It came within 5% of each, and the 1.618 and 4.236 rows sit outside the ±2% tolerance the framework applies elsewhere. More importantly, the 4.236 at 38,820 was not a terminal at all: the index passed through it and by July 2026 stood at 53,289, some 37% above. A level that price approaches and then leaves far behind has not been validated by the approach.
What can fairly be said is narrower. Over thirteen years spanning algorithmic trading, central-bank intervention and a pandemic, four extension levels drawn from a single swing sat near four significant highs, and the 2.618 sat within 126 Dow points of one of them. That is a striking coincidence if it is one. It is a single anchor, and four observations, and it proves nothing on its own.
Let's test a more specific hypothesis: Do extension ratios calculated in one era retain predictive power in subsequent eras?
Market structure evolves through "anchors"—major capitulation lows that reset the geometric template. We identified three primary anchors:
Each anchor occurred in a different technological and regulatory era. If market structure is mechanistically determined (dependent on specific trading infrastructure), then extensions calculated from the 1932 anchor should fail once we reach the algorithmic trading era. The technology is too different; the anchor is too old; the relevance should decay.
Null hypothesis: Predictive accuracy should degrade as temporal/technological distance from anchor increases.
Let's test this:
| Anchor Year | Test Period | Years Elapsed | Tech Eras Crossed | Accuracy (±5%) |
|---|---|---|---|---|
| 1932 | 1932-1942 | 0-10 | 0 | 100% |
| 1932 | 1942-1966 | 10-34 | 1 (telephone era) | 80% |
| 1974 | 1974-2000 | 0-26 | 1 (computer era) | 86% |
| 2009 | 2009-2024 | 0-15 | 0-1 (HFT/AI era) | 90% |
Result: No decay in accuracy is visible across temporal or technological distance. The 2009-2024 period, with the most advanced technology, scores no worse than the 1942-1966 period.
It would overstate this considerably to say the null hypothesis is rejected, which is what this page previously said. Each cell rests on between five and ten levels. A difference between 80% and 90% at those sample sizes is one event. More to the point, a study with cells this small has very little power to detect decay even if decay is occurring: the absence of a visible trend across four coarse buckets is weak evidence that no trend exists. What the table supports is the narrower statement that no decay is apparent in the cases examined, which is a reason to keep testing the proposition rather than a reason to consider it settled.
The proposition is nonetheless worth taking seriously, because the cross-market work referenced in Section 3 tests something adjacent on far more data and finds the effect positive in all twenty-five indices examined, spanning markets that adopted electronic trading decades apart. That is a better argument for technological indifference than the table above, and it is the one we would put forward.
How do we explain this? If technology, news events, and participant behavior have completely changed, but the geometric structure remains constant, what is actually driving price movement?
"Technology is the substrate, not the structure. The telegraph, the telephone, the computer, the algorithm—these are merely the transmission mechanisms for an underlying geometric effect. The external variables (news, sentiment, events) are placeholders that fill predetermined structural roles. The math precedes the narrative."
Consider an analogy: ocean waves. The specific water molecules in a wave are constantly changing—water enters at the base, rises to the crest, falls back down, and new water takes its place. But the wave form itself is stable. It maintains shape, amplitude, and wavelength even as the constituent particles continuously change.
Market price action operates similarly:
Random walk theory assumes that the molecules ARE the wave—that price movement is the direct aggregate of information-driven decisions. Change the information arrival mechanism (telegraph → fiber optic), change the decision-makers (specialists → algorithms), and you should get different price behavior.
Our evidence points the other way: the wave form appears largely independent of the molecules. Components of the system can be replaced—the technology, the participants, the information infrastructure—and the geometric structure is still recognisable.
This is consistent with markets being organised at a level beneath the observable mechanisms. It is worth stating the alternative honestly, because it has not been ruled out: proportional structures of this kind can also arise from the ordinary statistics of trending, mean-reverting price series, in which case the ratios would be a property of how prices move rather than a constraint on where they may go. Distinguishing the two requires showing that the observed ratios are tighter than such a process produces. The cross-market dispersion result—0.285 against 0.450 for a matched control—is the strongest evidence we have on that question, and it favours the structural reading without closing the argument.
Here's the uncomfortable implication: the news narratives we use to explain price movements are post-hoc justifications, not causal drivers.
Consider these "explanations" for major market peaks:
These narratives are completely different. Different technologies, different political contexts, different "causes." Yet each of these highs formed within about 5% of a 4.236 extension of its governing swing. The structural position was comparable—not identical, since 5% of the Dow is a wide zone—while the narrative explanation was unique to each.
The 2022 case belongs in this list with a caveat that materially weakens it. Unlike 1929 and 2000, the January 2022 high was not a generational terminal. The market fell about a fifth, recovered, and by July 2026 traded near 53,289—roughly 37% above the very 4.236 extension the peak is here credited with respecting. Whatever happened at 36,953, the level did not hold. Including it alongside two genuine tops, and omitting the several occasions when price approached a 4.236 and simply continued, is selection after the fact of the kind this series warns against elsewhere.
What does the coincidence tell us? Less than the previous version of this page claimed. It said: markets were geometrically positioned to peak at those levels; the news events that "triggered" the peaks were simply the available placeholders. That is a claim about causation drawn from three observations, and the direction of causation is exactly what cannot be read off them. The defensible statement is that the narrative offered for a turn varies wildly while the structural position varies less, which suggests the narrative is not doing as much explanatory work as it appears to. It does not establish that geometry drives price, that news is epiphenomenal, or that the structure was "already encoded." Those remain conjectures, and this page is titled a hypothesis for that reason.
If the Placeholder Hypothesis is correct—if technology and news are merely the substrate for a deeper geometric structure—then most of conventional market theory requires revision:
EMH Claim: "Prices reflect all available information; patterns cannot persist because they would be arbitraged away."
Problem: The Fibonacci structure remains recognisable across 130 years and five technological regimes. If it were readily arbitrageable, one would expect high-frequency algorithms to have compressed it. The 2009-2024 era, at peak HFT dominance, scores no worse than earlier ones. It should be said that "no worse" is all the data support: the era-by-era differences rest on a handful of levels each and a 90% against an 86% is not a finding.
Implication: The structure is either not arbitrageable (requires longer time horizons than algorithmic systems operate on), or it represents a constraint that even informed participants cannot violate.
Behavioral Claim: "Market patterns emerge from psychological biases (herding, loss aversion, overconfidence)."
Problem: Participants in 1896 were wealthy individuals making decisions over days. Participants in 2024 include algorithms making decisions in microseconds. Yet both eras trace comparable Fibonacci proportions. The contrast is weaker than it first appears—algorithms are written by people, encode their authors' objectives and priors, and are switched off by people under stress—so "zero psychological bias," as this page previously put it, overstates the case considerably. What can be said is that the mechanism of decision has changed beyond recognition while the proportions have not.
Implication: Psychology may fill the structure while not creating it. This is an inference, and a competing one survives: that human psychology is stable enough across 130 years, and sufficiently embedded in the algorithms that now intermediate it, that the persistence of the proportions is exactly what behavioural finance would predict. We do not know how to distinguish these possibilities with the data available, and we prefer to say so.
Complexity Claim: "Markets are complex adaptive systems; emergent properties depend on agent interactions and feedback loops."
Problem: The agent types, interaction speeds, and feedback mechanisms have completely changed. Yet the emergent geometric properties remain constant.
Implication: Either there is a meta-constraint operating above the agent level (some universal property of price systems), or markets are not primarily "adaptive" in the complexity theory sense—they are following a predetermined template that agents collectively fill.
We propose an alternative framework: Structural Primacy
This model would account for:
We began with a paradox: How can the same mathematical framework govern markets across 130 years of radical technological transformation? Random walk theory says it shouldn't. Behavioral finance says participant psychology should matter. Complex systems theory says emergent properties should change when the underlying agents change.
The evidence points one way without settling the matter: the proportions change far less than everything around them.
From telegraph wires to fiber optics. From paper tickets to algorithmic microsecond execution. From 12 stocks to 30. From specialist pits to dark pools. Through depressions, world wars, nuclear crises, tech bubbles, housing crashes, and pandemics. The Fibonacci structure remains recognisable.
Whether that constitutes a law is not something this page has established, and the word was previously used here too freely. What has been established, on data neither hand-picked nor scored against a null of our own choosing, is that price terminates on the ladder about 75% of the time where chance would give 54%, and that the terminal ratios are more tightly dispersed than a matched control produces, in every one of twenty-five markets. That is a real regularity. It is not gravity.
The placeholder hypothesis proposes something stronger and less comfortable: that the news does not move markets so much as furnish them with explanations after the fact. Our evidence is consistent with this. It does not demonstrate it, and the alternative—that stable human dispositions, now encoded in the machines that trade on our behalf, generate stable proportions—accounts for the same observations without requiring anything to precede anything.
Different ages. Same math.
There are three readings and we can eliminate only one. It is not a coincidence; the cross-market results are too consistent for that. It may be that markets obey an organising principle we have not formalised, treating information and technology as variables within a fixed geometric frame. Or it may be that the proportions are a by-product of how trending, mean-reverting price series behave, in which case the ladder describes the shape of price rather than constraining it—and would still be useful to a trader, while meaning something entirely different.
We favour the organising-principle reading. The evidence is sufficient to make it worth pursuing, and insufficient to make it worth asserting. The distinction between those two states is the whole of intellectual honesty in this field, and the reader is entitled to see us observe it.