Discretionary versus systematic: an honest comparison

Almost every argument about discretionary versus systematic trading is conducted badly, because both camps caricature the other. The systematic camp claims discretion is superstition with a Bloomberg terminal. The discretionary camp claims systems are curve-fitted arithmetic that breaks the moment the world changes. Neither description survives contact with people who actually do the work.

The real distinction is narrow and precise, and it is worth getting right before you write a single line of testing code. A systematic trader commits the decision before seeing the instance. A discretionary trader commits the decision after. Everything else — how many indicators are involved, whether a computer executes the order, how complex the logic is — is decoration. A trader with a written checklist who still chooses which signals to take is discretionary. A trader with a one-line rule that fires automatically is systematic, however crude the rule.

Why the timing of the commitment is the whole thing

If you commit to a rule in advance, the rule becomes a fixed function from inputs to outputs. The same inputs produce the same output on Tuesday morning and on Friday afternoon after a bad week. That property is what makes a backtest meaningful at all: you can replay history through the function and the answer you get is the answer you would have got. A discretionary process cannot be replayed, because the function includes a state variable — your mood, your recent P&L, your sleep — that is not in the historical record.

This is the mechanism behind the systematic case, and it has nothing to do with computers being clever. Psychologists have a term for it: noise, meaning unwanted variability in judgements that should be identical. Kahneman, Sibony and Sunstein's book Noise documents this across underwriting, medicine and sentencing: experts asked to evaluate the same case twice, weeks apart, frequently disagree with themselves. The older literature is starker still. Paul Meehl's 1954 monograph Clinical versus Statistical Prediction compared expert clinical judgement against simple mechanical formulas, and Grove and Meehl's 1996 meta-analysis extended the comparison across many domains: the formula matched or beat the expert in the large majority of studies. Robyn Dawes went further in 1979 with "The robust beauty of improper linear models in decision making", showing that even a linear model with arbitrary equal weights, fed the variables the expert said mattered, tended to beat the expert.

A trading system is exactly that: an improper linear model of your own judgement, with the noise removed. That is a substantial gift, and it is available to anyone willing to write their reasoning down.

What discretion actually buys, when it works

The honest counter-argument is that a fixed function is only as good as the variables it can see, and markets routinely produce situations whose relevant variables were never in the dataset. A systematic equity book running short positions in March 2020 met this when regulators in France, Italy, Spain, Belgium, Austria and Greece imposed emergency prohibitions on establishing or increasing net short positions in listed shares: what changed was not the return distribution but which trades were legal, decided at a few days' notice by supervisors rather than by the market. A discretionary trader following the regulatory debate in those weeks had access to information no price-history model could encode.

Discretion earns its keep in three specific places. First, structural breaks — central bank regime changes, sanctions, delistings, exchange rule changes — where the historical distribution is simply the wrong reference class. Second, thin data problems, where the interesting event has happened four times and no statistical test can help you. Third, breadth of context, where a human integrates a filing, a conversation and a chart in a way that no feasible feature set captures.

Note what these have in common: they are all cases where the sample is too small or too unrepresentative for a rule to be estimated. That is the correct dividing line. Where you have hundreds of comparable observations, the rule should win. Where you have four, judgement is all you have.

The arithmetic of overriding your own system

The hybrid — run a system, but skip trades that "look wrong" — is the most common arrangement and the least examined. It is worth doing the arithmetic, because the range of outcomes is enormous and invisible without records.

Define R as the amount of capital you risk on one trade, so a +2R trade makes twice what a losing trade loses. Suppose a rule produces 200 trades a year with a 40% win rate, average win 2R and average loss 1R. Expectancy per trade is (0.40 × 2R) − (0.60 × 1R) = 0.8R − 0.6R = 0.2R, so the full system earns 200 × 0.2R = 40R a year.

Now suppose you override it and skip a quarter of the signals — 50 trades — because they look wrong.

If your overrides are pure noise, the 150 remaining trades keep the same 0.2R expectancy and you earn 30R instead of 40R. You have paid 10R for the privilege of having opinions.

If your overrides are genuinely skilled and the 50 skipped trades would have averaged −0.3R each, those trades were dragging −15R out of the total. The 150 you took therefore contributed 40R + 15R = 55R, or 0.367R each. Your intervention lifted the year from 40R to 55R — a 37.5% improvement.

If your overrides are anti-skilled and you skipped the 50 best trades, which averaged +0.8R, those alone accounted for 40R. The remaining 150 contributed exactly zero, and your year is flat.

Same system, same override rate, outcomes from 0R to 55R. The uncomfortable part is that you cannot tell which trader you are unless you log the trades you did not take and mark them to market as if you had. Almost nobody does this, which is why almost nobody knows whether their discretion adds or destroys value. If you intend to run a hybrid, that log is not optional bookkeeping; it is the only instrument that measures the thing you are betting on.

The hidden discretion inside every system

Systematic traders should resist smugness here. Every system contains a large amount of judgement; it has simply been moved upstream into the research process. Someone chose the universe, the lookback windows, the exit logic, the rebalancing frequency, the cost assumptions and — crucially — which of the forty variants tried in testing was the one that got deployed. That last choice is discretion applied to a sample, which is worse than discretion applied to a live market, because it leaves no trace in the equity curve and it flatters you systematically. The rest of this course is largely about controlling it: the lessons on overfitting, on in-sample versus out-of-sample design, and on look-ahead bias all exist because the researcher's freedom is the main threat to a system's honesty.

So the correct statement is not "systems remove judgement". It is: systems relocate judgement to a place where it can be audited, and the audit is your responsibility.

Failure modes are different, not smaller

Choose with the failure modes in view, because they are not comparable in kind.

Discretionary trading fails through drift and noise: the process quietly changes shape after a drawdown, position sizing becomes emotional, and the trader cannot reconstruct what they were doing six months ago. It also fails through unfalsifiability — a process that is never fully specified can never be shown to have stopped working.

Systematic trading fails through overfitting and regime dependence: a rule that described the past beautifully because it was tuned to it, or a rule that was genuinely real in one monetary regime and is not in the next. It also fails through operational fragility — a data feed changes format, a symbol is renamed, and the machine keeps trading on garbage without hesitation, which no human would do.

Systems also have a capacity and attention profile that suits people with day jobs. A rule evaluated on daily closes takes minutes a day to run and does not require you to be at a screen when the SIX Swiss Exchange opens. That is a legitimate and underrated reason for a professional in another field to prefer the systematic route, quite apart from any claim about edge.

The Swiss overlay: how institutions are made to pre-commit

If the argument for writing the decision down in advance sounds like a private preference, it is worth noting that Swiss law imposes exactly that structure on the largest pool of investment capital in the country. Occupational pension assets are governed by the BVV2 ordinance, which does two things at once. It sets ceilings by asset category — limits on equities, on real estate, on alternative investments, and on foreign-currency exposure left unhedged — and it requires the fund's governing body to adopt a written investment reglement stating the strategy, the permitted ranges and who may deviate from them.

The interesting part is what BVV2 does with discretion. The category limits are not absolute prohibitions: a fund may exceed them, but only if it can justify the extension in its annual accounts, on the basis of a documented assessment that the arrangement remains prudent. In other words, the regulator did not ban judgement. It required that judgement be committed to paper, attributed, and defended after the fact against a benchmark that was written down first. That is the discretionary override of this lesson, formalised — and the reason it is formalised is that unrecorded overrides cannot be evaluated, by a supervisor or by you.

The lesson for a private trader is not that you need an investment reglement. It is that the one institution with genuine accountability for other people's retirement money reached the same conclusion this lesson does from a different direction: discretion is permissible, and it is only permissible when the baseline it departs from and the reason for the departure both exist in writing before anyone knows how it turned out.

What to take away

• The defining difference is when the decision is committed — in advance for a system, on the instance for discretion — and only an advance commitment can be honestly backtested.
• The measurable advantage of a system is the removal of noise, the unwanted variability in judgements that should have been identical; this is a well-documented finding across professions, not a claim unique to markets.
• Discretion has a genuine domain: structural breaks, events with too few observations to estimate, and context that no feasible dataset encodes.
• If you override your system, the same 25% override rate can improve your year by nearly 40% or wipe it out entirely, and the only way to know which is to log and mark the trades you skipped.
• Systems do not eliminate judgement; they move it into the research process, where it must be audited — which is the subject of most of the lessons that follow.
• Swiss occupational pension funds are required under BVV2 to run from a written investment reglement and to justify in their annual accounts any departure from the category limits — institutional confirmation that discretion is only defensible when the baseline and the reason for leaving it were both written down first.

Sources

• Daniel Kahneman, Olivier Sibony and Cass Sunstein, Noise: A Flaw in Human Judgment (2021) — on variability in expert judgement and the case for rules.
• Paul E. Meehl, Clinical versus Statistical Prediction: A Theoretical Analysis and a Review of the Evidence (1954); and William Grove and Paul Meehl, "Comparative Efficiency of Informal and Formal Prediction Procedures", Psychology, Public Policy, and Law (1996).
• Robyn M. Dawes, "The Robust Beauty of Improper Linear Models in Decision Making", American Psychologist (1979).
• Robert Carver, Systematic Trading: A Unique New Method for Designing Trading and Investing Systems (2015) — a careful practitioner account of where rules help and where they do not.
• Andrew W. Lo, Adaptive Markets: Financial Evolution at the Speed of Thought (2017) — on why edges are regime-dependent rather than permanent.
• Jack D. Schwager, Market Wizards (1989) and Hedge Fund Market Wizards (2012) — interview evidence on how discretionary and systematic practitioners describe their own processes.
• Ordinance on Occupational Retirement, Survivors' and Disability Pension Plans (BVV2 / OPP2), provisions on the written investment reglement, the category investment limits and the conditions for extending them.
• ESMA public statements of March 2020 on short selling, and the temporary short-selling prohibitions issued that month by the French AMF, Italian CONSOB, Spanish CNMV, Belgian FSMA, Austrian FMA and Greek HCMC — the reference case for a rule change that no price history encodes.