A hypothetical researcher tests 50 variations of a breakout rule and keeps the best result. The selected chart looks convincing, but it contains a hidden advantage: it won a contest against 49 alternatives on the same history. David Harding’s role as founder of Winton provides a useful setting for examining that problem.
Winton’s biography identifies Harding as its founder and chairman and dates the firm’s establishment to 1997. Its history describes an initial focus on futures trend following and later development across additional data, assets and strategies. Reducing the firm to one breakout formula would miss that evolution.
The examples below are independent hypothetical research exercises, separate from any Winton model or Harding trade. They make a research claim inspectable: what was tested, what information was available, what choices were made and what remains uncertain after a favourable result.
David Harding and the cost of choosing the best backtest
Suppose 50 candidate rules use different combinations of lookback periods and exit distances. The researcher reports only the version with the highest historical return. Even if every candidate is fully mechanical, the selection introduces discretion at the research stage.
The result now reflects both the market history and the search across alternatives. A rule that happened to suit a few unusual trades can rise to the top. Repeating the search on the same data doesn’t provide fresh evidence.
I would preserve the count of 50 trials beside the selected result. It changes how much confidence the result deserves compared with a single hypothesis specified before the data was examined. Keeping the losing candidates also reveals whether the apparent success is isolated or shared by nearby settings.
A research log can record the candidate, the reason for trying it and the date of the test. It can distinguish planned experiments from ideas prompted by an earlier result. That makes the actual search visible, including ideas abandoned after disappointing results.
A breakout rule needs a precise information boundary
Consider an illustrative rule that acts when the completed daily close exceeds the highest high of the preceding 20 sessions. If that prior high is 60.00 and today’s close is 60.20, the condition qualifies. A high of 60.40 with a close of 59.80 does not.
The word “preceding” matters. Including today’s high in the threshold changes the condition. A close ordinarily cannot exceed its own session’s high, so a seemingly minor implementation choice can eliminate the intended signals.
The Donchian Channels guide gives context for a lookback-based reference. The study still needs an entry convention. Entering at the next open uses a price that follows the completed signal; assuming an exact fill at the final close requires a feasible execution arrangement.
A gap from 60.20 to 61.10 changes the available entry by 0.90. Ignoring it can distort both returns and the distance to a protective reference. A reproducible signal definition is only one part of an executable strategy.
Reserve observations the search has not already used
A hypothetical project might use an earlier period to develop the rule and reserve a later period for evaluation. The important distinction is whether decisions were made using the reserved results. Once the researcher repeatedly adjusts the model after inspecting them, that period becomes part of development.
For illustration, suppose an earlier ten-year period produces a positive result but the next three years lose money. The new evidence weakens the case. It doesn’t automatically prove the model is permanently invalid, because sampling variation and changing conditions remain possible explanations.
A useful report preserves the original rule and the disappointing result before proposing a revision. If the exit changes from a ten-session reference to a five-session reference, the revision is a new candidate with a new evidential burden.
The Ed Seykota profile provides a related discussion of systematic practice. Consistency makes a rule’s behaviour easier to inspect. It cannot substitute for evidence that the rule’s behaviour is economically useful.
Compare exposures before adding markets
Two hypothetical futures contracts each move by one price point. In one contract, a point is worth 10 monetary units; in the other, it is worth 100. Holding one of each gives very different sensitivity to an identical numerical move.
Suppose their daily movement estimates are 2 points and 0.5 points. Multiplying by the contract values gives estimated daily monetary movements of 20 and 50 per contract. A simple equalisation exercise would require different quantities.
I find the 20-versus-50 comparison more informative than the two contract counts. It shows why an instrument list alone cannot describe diversification. The exposure depends on contract specifications, quantities, currency conversion and the way the markets move together.
Historical volatility offers one estimate of past variability. It isn’t a ceiling on future movement. A sudden regime change can make yesterday’s apparently balanced quantities produce uneven losses today.
A smooth average can hide a difficult sequence
A hypothetical strategy earns 20R across 100 trades. That total says little about the order of the outcomes. Ten consecutive losses near the beginning create a different experience from the same losses scattered through the record.
A review can retain the equity path, the longest losing sequence, the largest peak-to-trough decline and the time required to recover. Each adds information that the average return leaves out. None supplies a maximum future drawdown.
The Ulcer Index is one way to summarise drawdown observations over a chosen window. Its result depends on the sampled path and frequency. A monthly series can conceal movements visible in daily data.
For the fictional model, a researcher might compare the original trade sequence with alternative orderings as a sensitivity exercise. That can illustrate path dependence, but it doesn’t recreate all real market relationships. Shuffling trades can remove the clustering the strategy would actually encounter.
Changing data can change the apparent discovery
Suppose a stock database contains only companies that remain listed today. A historical strategy tested on that set excludes companies that disappeared. The test’s universe is then shaped by future survival.
A hypothetical screen selecting the strongest 20 stocks in a past year needs the universe that was available at that time, including later failures. Otherwise the test may appear better because some of its possible losers were removed before the researcher began.
Futures data present different construction questions. A continuous price series may adjust past prices when contracts roll. The research needs to distinguish the series used for the signal from the actual contracts and prices used to calculate tradable returns.
These details deserve a place in the specification rather than a footnote added after a promising result. A small model with clear data assumptions can be easier to assess than a complicated model whose historical inputs cannot be reconstructed.
Monitor a live process without rewriting every disappointment
Imagine a hypothetical simulation expects a complete-trade cost of 0.10R, while the first 30 observed trades average 0.18R. The discrepancy is measurable even if the strategy’s gross result is favourable.
A review can examine spreads, order sizes, timing and missed fills. It can also check whether the simulation and observed record classify the same events. A mismatch in timestamps can look like poor execution when the two records are actually describing different opportunities.
Thirty trades provide limited evidence about long-run returns, but they may be enough to expose a persistent accounting error. Different questions need different amounts of evidence. Waiting for a large performance sample is unnecessary when a contract multiplier is demonstrably wrong.
The corrective action depends on the finding. Fixing a calculation error restores the intended model. Changing the signal after losses creates a revised model. Recording that distinction keeps maintenance from quietly becoming another search for a favourable backtest.
Related books explain methods without revealing Winton’s models
Following the Trend, by Andreas F. Clenow, offers related study of systematic futures trading. Trend Following, by Michael Covel, supplies broader treatment of the approach. Neither title is authored by Harding or an authorised disclosure of Winton’s internal models.
A practical reading exercise is to write a complete specification for the 20-session breakout example, including the lookback convention, next-session fill and contract multiplier. Gaps in that specification show what still needs to be resolved before a test can be interpreted.
The Man Who Solved the Market, by Gregory Zuckerman, concerns Jim Simons and the quant revolution. It can provide comparison with a different figure in quantitative investing, but it isn’t a Harding biography. Keeping those identities clear prevents a compelling story from being used as evidence for the wrong person.
Disclosure: As an Amazon Associate, I earn from qualifying purchases.
Following the Trend is linked in the second-edition hardcover: (US) Following the Trend, Second Edition, (UK) Following the Trend, Second Edition, (DE) Following the Trend, Second Edition.
Trend Following is linked in the 2009 updated paperback edition: (US) Trend Following, 2009 Updated Edition, (DE) Trend Following, 2009 Updated Edition.
The Man Who Solved the Market, by Gregory Zuckerman, is linked in paperback as comparative reading about Jim Simons: (US) The Man Who Solved the Market.
Keep the rejected experiments beside the winner
A selected model becomes easier to assess when its 49 rejected alternatives remain in the research log. The winning chart then appears in the context that produced it. That context is part of the evidence, and it deserves to survive the moment a result becomes attractive.
Learn the pattern. Ride the trend. Keep the gains.
Educational content only. Not investment advice. Trading involves risk. You are responsible for your decisions.
Get the free Market Wisdom e-book
Join Trends and Breakouts — historical winners, breakout studies, and risk lessons. No spam, unsubscribe anytime.