Fibonacci retracement: the ratios, the anchor problem, and the honest limits

Before any argument about whether the levels mean anything, it is worth being exact about what the software is doing, because almost nothing is written down plainly. The whole tool is one line: level = high minus ratio times (high minus low). You give it two prices. It multiplies the distance between them by a constant and subtracts the result from the top. That is all. There is no model, no fitting, no data beyond the two numbers you nominated, and nothing your platform computes here that you could not do on the back of an envelope in about fifteen seconds.

Run it. Take an illustrative up-swing on an index from a swing low of 21,800 to a swing high of 24,400, so the range is 2,600 points. The 61.8 percent level is 24,400 minus 0.618 times 2,600, which is 24,400 minus 1,606.8, or 22,793. The 38.2 percent level is 24,400 minus 993.2, or 23,407. Every other level is the same operation with a different constant, and the table below is that formula run five times and nothing else.

Now look at what the formula takes as input. Two numbers: a high and a low. Everything else is a constant that never changes. So every level on your chart is a function of a choice you made about where a swing begins and ends, and the formula has no way to check that choice. It will compute a confident, precise, five-decimal grid from any two prices you nominate, including a pair you chose badly, including a pair you chose because the resulting grid looked better. All of the tool's precision sits downstream of an act of judgement that has no precision at all. Hold on to that, because it is where this guide ends up.

The sequence starts 1, 1, 2, 3, 5, 8, 13, 21, 34, 55, 89, 144 and grows by one rule: each term is the sum of the two before it. Divide a term by the next one and the answer does not march politely toward its destination. It overshoots and undershoots and closes in: 1 divided by 2 is 0.5, then 0.667, then 0.6, then 0.625, and by 34 divided by 55 you are at 0.61818, and by 144 divided by 233 at 0.61803. The limit is 0.6180339887, the reciprocal of the golden ratio, 1.618034. That convergence is the real content of the idea: the ratio belongs to the sequence itself, not to any particular pair of numbers in it.

The rest of the family is that single number rearranged, and the arithmetic is prettier than most trading guides let on. The golden ratio satisfies the relation phi squared equals phi plus one. Divide through by phi squared and you get 1 divided by phi, plus 1 divided by phi squared, equals 1. So 0.618034 plus 0.381966 is exactly 1. The 38.2 percent level is not an independent discovery about markets or anything else: it is 1 minus 0.618, the same fact counted from the other end of the swing. Cube the number and you get 0.236068, the 23.6 percent level. Three levels, one constant, no approximation.

Then there are the passengers. The 50 percent line is not in the sequence anywhere, and its own standard reference says so: StockCharts states that the 50 percent retracement is not based on a Fibonacci number and that it stems instead from Dow theory's assertion that the averages often retrace half their prior move. It is a much older and much simpler idea wearing a borrowed coat, and it survives inside the tool because a halfway mark is genuinely useful and because it happens to sit conveniently between two of the real ratios. Your platform draws it in the same colour, in the same style, at the same time as the others, and so it is inherited as a Fibonacci level by almost everyone who uses one.

The 78.6 percent line is stranger, and more revealing. It is the square root of 0.618. A square root of a ratio is not a ratio between terms of the sequence, so it is a different kind of object with a Fibonacci label on it. Ask why it exists and the answer is uncomfortable: there was a conspicuous gap between the 61.8 percent line and the swing low, and taking a square root landed a new line inside it. Harmonic traders subsequently took the square root again, arriving at 0.886, which fills the next gap along. The pattern is the point. Each addition to the set is whichever operation places a fresh line in the widest remaining space.

None of this is an attack on the sequence. The mathematics is exact, it is genuinely elegant, and every number above is computed rather than asserted. The attack is on a step that nobody takes: the move from a property of a recurrence relation to a claim about where a bank's share price stops falling. That step is where the whole edifice rests, and it is the one part of the subject nobody writes down.

Between the statement "0.618 is the limit of the ratio of successive terms" and the statement "price reverses at 61.8 percent of a swing" there is supposed to be an argument. Go and look for it. It is not in the trading literature, it is not in the platform documentation, and it is not in the courses. The levels are asserted, drawn, and then defended by example. This is not a small gap in an otherwise sound case. It is the entire case, missing.

The usual gesture is toward nature: sunflower heads, nautilus shells, spiral galaxies, the proportions of the human body. There are two problems with it, and the second is much worse than the first. The first is that the claims are far weaker than advertised. George Markowsky went through the standard examples in the College Mathematics Journal in 1992 and found the mathematical properties stated correctly but much of what is claimed for the golden ratio in art, architecture and aesthetics to be false or seriously misleading. The Parthenon, the golden rectangle of the classical canon, the proportions of the body: largely retrofitted, measured generously, or simply wrong.

The second problem is the one that settles it. Suppose every natural claim were true. It would still carry no weight here. Phyllotaxis is a packing result: a growing plant that places each new primordium at a fixed angle from the last achieves an efficient, non-overlapping arrangement, and the golden angle is the one choice that never repeats itself. That is a fact about growth under a geometric constraint. A share price is a clearing price between two people's expectations about the future. The two phenomena have nothing whatsoever in common except a number, and a number is not a mechanism. Nobody has ever proposed a channel by which the first could reach the second, which is not a matter of the argument being weak. There is no argument.

A sunflower and a share price have nothing in common but a number. A number is not a mechanism.

It is worth being precise about the burden here, because the honest claim is narrow and the dishonest one is broad. Nobody needs to prove that the levels do not work. That is not how this runs. A tool that claims price responds to one specific ratio owes a reason, and no reason has been offered. What has been offered instead, for decades, is the picture: a chart, after the event, with a line on it.

There is exactly one explanation that does not require mysticism, and it deserves to be put at its strongest before anything is done to it. Call it coordination. In 1960 Thomas Schelling posed a puzzle in The Strategy of Conflict: you must meet a stranger in New York tomorrow, you cannot communicate, and you both know only that the other is trying to find you. People converge, overwhelmingly, on Grand Central at noon. Nothing about that clock is special, and no payoff makes it correct. It is salient, and each person picks it because they expect everyone else to. Schelling called these focal points, and they are one of the few genuinely load-bearing ideas in the social sciences.

A Fibonacci level is a candidate focal point, and a strong one, for an entirely boring reason: it is the platform default. Open any charting package, drag the tool across the obvious swing, and it draws 61.8 percent at precisely the same price for everyone who chose the same swing. So the chain runs like this. The platform draws the line. Many traders see the same line. Their buy limits, their stops and their targets land on one price. That cluster of resting orders is real supply and demand at a price, not a theory about one. And price pauses there. Every link in that chain is an observable step, there is nothing magical in it, and it is exactly the mechanism behind a pivot point or a round number: the level matters because a crowd is looking at it, and for no other reason. What you are ultimately looking at is a supply and demand imbalance that a shared drawing convention helped to assemble.

Now read what that explanation costs, because it is usually offered as a defence of the tool and it is nothing of the kind. Go back through the chain and try to find the golden ratio doing any work in it. It is not there. Phi's only contribution was to be the number the software draws by default. Any number would serve, provided the crowd has agreed on it: yesterday's high, a round 24,000, the pivot, the opening price. The coordination account, taken seriously and followed to its end, is not a defence of the golden ratio at all. It is a demonstration that the golden ratio is inert and the agreement is the active ingredient. The strongest case anybody can make for Fibonacci is simultaneously a proof that Fibonacci is not what is doing the work.

And coherent is not the same as demonstrated. A chain of plausible links is a hypothesis, not a finding, and this one has been looked at. The most careful test we can point to is Nikhil Gupta's 2011 honours project at Macalester, which examined Fibonacci retracements across six major currency pairs on tick-by-tick data from 2003 to 2008, over horizons from intraday to monthly. It found the levels lacked economic significance in every period and at every horizon, and it read that null result as evidence against the self-fulfilling account rather than for it. That is currency majors rather than an Indian equity index, and it does not close the question here. But it is the only real evidence in the room, it looked precisely where the mechanism should have shown up, and it found nothing. Pages that present the self-fulfilling story as established fact, and there are many, are citing a hypothesis and calling it a result.

Return to the formula, because the fatal problem was visible in it from the first section. It needs a high and a low. The chart does not hand you those; you pick them. And there is no rule for picking them. Worse than no rule: the literature offers two standard instructions that routinely contradict each other. Anchor to the most significant swing. Anchor to the most recent impulse leg. Both are taught, both are defensible, both appear in reputable material, and on any chart with more than one leg in it they disagree.

Watch them disagree. An index advances from 21,800 in an irregular run up to 23,400, corrects to a higher low at 22,900, then runs on to a swing high of 24,400. Analyst A measures the whole advance: 21,800 to 24,400, a swing of 2,600 points. Analyst B measures the final impulse leg: 22,900 to 24,400, a swing of 1,500 points. Neither of them has made a mistake. Same chart, same formula, same discipline, different input.

The result is not a small discrepancy. A's golden-ratio line sits at 22,793. B's sits at 23,473. They are 680 points apart, which on this chart is 26 percent of the entire swing being measured, and not one of A's five levels coincides with any of B's. Two competent analysts, looking at one screen, produce two complete and internally consistent grids that share nothing at all. There is no procedure either of them can run to settle which grid is the real one, because there is no fact of the matter about which swing the market is measuring. The market is not measuring a swing. They are.

Then price does what price does. The pullback ends at 23,450 and turns back up, and now watch both analysts be right. For A, that low is 43 points above the 38.2 percent line: a shallow, healthy pullback that respected 38.2 percent and confirmed the trend. For B, the very same low is 23 points below the 61.8 percent line: the textbook golden-ratio retracement, respected almost to the tick. Both will screenshot it. Both are telling the truth about their own grid. And they cannot both be evidence that price obeys the golden ratio, because one of them is saying that the golden ratio was the wrong line and 38.2 percent was where the work happened.

Now put this next to the previous section, because the two problems are not neighbours. They are the same problem. Coordination was the only thing holding the tool up, and coordination requires agreement. The load-bearing link in that chain was "many traders pick the same obvious swing." If the swing really is obvious, the crowd agrees, the orders cluster at one price, and the mechanism runs. If the swing is ambiguous, there is no agreement, so there is no cluster, so there is no mechanism, so there is nothing there at all. The anchor problem does not sit beside the coordination story as a separate caveat. It dissolves it, and it dissolves it exactly where you were hoping for help.

Which leaves the tool somewhere unflattering. It can only work where the swing is so obvious that everybody draws the same one, and where a swing is that obvious you did not need a tool: you were already looking at a clean structure with an unmistakable high and low, and you could have marked those two prices yourself. It fails precisely where charts are hard, where the structure is ambiguous and several readings compete and you genuinely wanted assistance, because ambiguity is the exact condition under which a crowd cannot coordinate. The tool is strongest where it is redundant, and weakest where it is needed.

The sophisticated defence, and the one experienced traders reach for, is that nobody trades a bare Fibonacci level. You use confluence. When 61.8 percent lands on a prior swing low, or a round number, or a moving average a lot of people watch, the zone is stronger than either factor alone. As a description of practice that is true, and the instinct behind it is correct. But look carefully at what it concedes, and then look at where it puts the credit.

The concession comes first, and it is larger than it appears. The moment you say a Fibonacci level needs an independent structure sitting next to it before you will act on it, you have conceded that the ratio is not sufficient. Fine. The interesting question is then whether it is necessary, and confluence is invariably presented as though the structure confirms the ratio. Run the counterfactual instead, which is the standard way to find out which part of a compound explanation is load-bearing: delete one factor at a time and see what survives.

Take the zone: an old shelf at a price, and a 61.8 percent line landing on the same price. Delete the Fibonacci line. What changes? The shelf is still there. The traders who remember it are still there. Their orders are still resting on it. Essentially nothing changes, and the zone holds for exactly the reasons it held before. Now delete the shelf instead and leave the Fibonacci line by itself. What remains is a thin residue of orders from whoever drew that particular anchor and happened to agree with your reading of it, which as the last section established is a much smaller and much less reliable group than the tool's popularity suggests. The zone is a different object entirely.

So the shelf is load-bearing and the ratio is a passenger. A Fibonacci level sitting on an old low is a renaming of the old low. It arrived at an address that was already occupied, and it took the credit for the neighbourhood. This is not an argument that confluence is useless, and it is not an argument against marking these zones. It is an argument that confluence as it is normally taught has the attribution backwards: the structure is not confirming the ratio, the ratio is borrowing the structure's authority and being graded on the loan.

There is a nastier consequence, and the previous section set it up. Two defensible anchors give you ten lines on one chart. Ten arbitrary lines across a pullback band will land near each other by arithmetic alone, with no help from the market. On the illustrative chart, A's 23.6 percent level at 23,786 sits 41 points from B's 38.2 percent level at 23,827. A trader who has drawn both grids will look at that pair and call it confluence. It is not confluence. It is two grids you drew yourself agreeing with each other, which they were always going to do somewhere. Confluence means something only when the agreeing factors are independent, and two Fibonacci grids on one chart are the least independent objects on it.

Everything above is argument, and argument is cheap. The honest move is to state what would change our minds and then go and look. This is the section every Fibonacci page skips, and the reason it gets skipped becomes obvious the moment you try to write the test down: specifying it carefully is enough, by itself, to show why the tool has survived so long without one.

Five things have to be fixed before you look at any data, and each one corresponds to a way the question is usually dodged. The anchor rule has to be mechanical, or you are fitting rather than testing. The tolerance has to be a number, because "near the level" is not a claim. "Held" has to be defined in advance. The sample has to include every occurrence the rule generates, misses included, since counting only the touches is not evidence but a scrapbook. And there has to be a control.

The control is the whole thing, and it is the piece nobody runs. Suppose you do the work and find that 61.8 percent holds some fraction of the time. That number, standing alone, means precisely nothing. The question was never whether the level holds. It is whether it holds more often than a line drawn at an arbitrary depth on the same swings, in the same instrument, over the same period. Without that comparison you have measured how often price pauses somewhere, which is a property of price, and then attributed it to the ratio. A base rate without a control group is not evidence, and the discipline of counting both arms honestly is the same discipline that separates a real backtest from a flattering one.

The tolerance is where it rots completely, and this is worth computing rather than asserting. Allow a level a tolerance of 0.4 percent of price, about 98 points on a 24,400 index, which is not a generous setting by the standards of how the tool is actually discussed. Draw both defensible grids from the previous section. Between 23,002 and 23,925 there is now a continuous 923-point stretch of chart in which every single possible price sits within tolerance of some Fibonacci level, with no gap anywhere in it, and across the whole band the two grids cover 80 percent of the ground. A pullback ending anywhere in that stretch confirms the tool. So does one ending almost anywhere else.

Note carefully what that last figure is and is not. It is not a finding about markets. It is arithmetic about two grids, and that is exactly why it is damning: no market behaviour was required to produce it. A rule that cannot be missed cannot be evidence of anything, and a practitioner who draws two grids has built a rule that cannot be missed. This is the multiple-testing problem wearing a chart. Bailey, Borwein, Lopez de Prado and Zhu put the general case in the Notices of the American Mathematical Society in 2014: the more configurations you try, the greater the probability that your best-looking result is an artefact of the search rather than a property of the world. A Fibonacci practitioner runs that search continuously and informally, with two or more anchors, five or six levels on each, a tolerance dial, and the freedom to set every one of them after seeing how it turned out.

Strip all of it back and ask what survives, because something does. A retracement grid is a fast, standard way to pre-mark a handful of prices before a session, and a shared vocabulary for describing how deep a pullback is. Both are real. "The pullback is at 38.2 percent" communicates a depth to another trader faster and more precisely than "it gave back a bit over a third of the move." And if a large number of participants are watching a line, that line is worth having on your chart, not because it has any power but because they do. That is the honest case for the tool, and it is the whole of it.

What does not survive is the inference. The grid gives you prices to watch. It never gives you a reason, and the distinction between those two things is the distinction between a drawing aid and an analysis.

Which gives the rule this whole guide has been walking toward, and it is short enough to apply at the moment it matters. A Fibonacci level may put a price on your list. It may never be the reason you took the trade. The reason has to be something that would still be standing if you had never opened the tool: a structure with a history, a level that a lot of people remember for reasons that have nothing to do with a ratio, a plan you wrote before the pullback started. Delete the grid from your reasoning and see what is left. If the trade still stands, take it. If it does not, you never had a trade. You had a drawing.

A drawing aid can put a price on your list. It can never be the reason you took the trade.

The habit generalises, and that is the real reason this page is worth its length. Every tool on a chart deserves the same interrogation: what is the proposed mechanism, what would falsify it, what is doing the work when it appears to succeed, and what would the honest control look like. Most tools in the standard technical toolkit answer those questions better than Fibonacci does, and a few answer them worse. Working out which is which, rather than accepting a set of lines because a platform draws them by default, is most of what separates reading a chart from decorating one. Structure first, and instruments only where they survive the question, is what the method we teach is built around.

• The derivation, and the 50 percent admission. StockCharts ChartSchool derives the 23.6, 38.2 and 61.8 percent levels by dividing a term of the sequence by the term one, two or three places ahead, and states directly that "the 50% retracement is not based on a Fibonacci number" but "stems from Dow Theory's assertion that the Averages often retrace half their prior move." chartschool.stockcharts.com
• The evidence, such as it is. Nikhil Gupta, Fibonacci Retracements and Self-Fulfilling Prophecy (Macalester College Economics Honors Projects, 2011), examined six major currency pairs on tick-by-tick data from 2003 to 2008 across horizons from intraday to monthly, found the retracements to lack economic significance in all periods and at all horizons, and read that result as running against the self-fulfilling hypothesis rather than supporting it. It is FX rather than Indian equities, and it is the reason this guide treats the coordination account as an unproven hypothesis. digitalcommons.macalester.edu
• The golden ratio outside markets. George Markowsky, Misconceptions about the Golden Ratio, The College Mathematics Journal, volume 23, number 1 (1992), pages 2 to 19, finds the mathematical properties generally stated correctly while much of what is claimed for the ratio in art, architecture, literature and aesthetics is false or seriously misleading. The nuance matters here: the mathematics is sound, the applications are where the trouble starts. tandfonline.com
• Focal points. Thomas C. Schelling, The Strategy of Conflict (Harvard University Press, 1960), introduced the idea that people coordinate without communicating by converging on whatever option is salient to all of them, each choosing it because they expect the others to. This is the mechanism, and the only one, under which a drawn level could matter.
• Searching until something looks good. David H. Bailey, Jonathan Borwein, Marcos Lopez de Prado and Qiji Jim Zhu, Pseudo-Mathematics and Financial Charlatanism: The Effects of Backtest Overfitting on Out-of-Sample Performance, Notices of the American Mathematical Society, volume 61, number 5 (2014), show that the more configurations are tried, the greater the probability the best-looking result is an artefact of the search. papers.ssrn.com
• On the arithmetic in this guide. Every ratio, level and coverage figure here is computed rather than quoted, and can be checked with a calculator. The identity that 1 divided by phi plus 1 divided by phi squared equals exactly 1 follows directly from phi squared equals phi plus one. The chart path, the two anchors and the 923-point coverage stretch are illustrative constructions, authored to make a point about arithmetic, and are not observations of any market.