Risk management

Position size is calculated, it is not chosen

Plenty of beginners pick their size by feel, or worse, by conviction: large when they believe, small when they doubt. That is the exact opposite of a system, since conviction is the poorest predictor you have. It is actively harmful, too: the setups that inspire the most confidence are the ones that look most like setups you have already seen win, which is a fair definition of confirmation bias.

Size is calculated, and the calculation fits on one line. Three numbers go in, one comes out. The first is the share of your capital this trade is allowed to cost, one per cent for instance. The second is the distance to your stop, measured in points, pips or ticks depending on the instrument. The third is the value of one point for a single unit of position. The size you want is the first number divided by the product of the other two.

Run the arithmetic once and it will never leave you. Capital of £10,000, risk set at 1%, so £100 on this trade. Your stop sits 40 points from your entry, and one point is worth £2 per lot. Each lot therefore risks 40 × 2 = £80, and your size is 100 ÷ 80 = 1.25 lots. Change one input: the same trade with a stop 100 points away puts the risk per lot at £200, and your size drops to 0.5 lots. The amount at risk has not moved by a penny.

What comes next always surprises people, and it is the heart of the matter: your size changes on every trade, and that is correct. A distant stop forces a small position, a close stop allows a large one, for an identical risk in money. A fixed size does exactly the reverse: it lets the volatility of the day decide how much you lose. On the morning the market turns jumpy and stops have to widen, your loss doubles without you deciding anything at all.

The twin mistake is to widen the stop afterwards so the size you wanted still fits. The stop was not moved by the market, it was moved by you, and the one number you genuinely controlled has just escaped. ⛔ If the calculated size looks laughable, the calculation is never the problem. The problem is that this trade, with this stop, is too expensive for this account, and turning it down is a sound decision rather than a missed opportunity.

A stop is a price that invalidates, not a sum you agree to lose

A stop is not an amount you agree to lose. It is a price that invalidates your reading. If that price prints, the reason the position exists is gone, and staying means holding a bet you yourself wrote off as wrong before you took it.

The question to ask before every entry fits in one sentence: which price, if it trades, proves I was wrong? If you cannot answer it, you do not have a trade, you have an urge. That question has a rare quality: it is asked while you are still calm, and it produces an answer another person can check on your chart. A feeling that the market looks tired cannot.

So the stop goes where the chart says it goes, behind the low that acted as support, beyond the zone price left, and not where your account would prefer. Many people run the reasoning backwards: they decide to risk £100, derive a distance of twenty points from it, and place the stop twenty points away. That is picking the answer before asking the question, and the market notices before you do.

Moving it the wrong way turns a planned loss into an unknown one. That single gesture empties accounts, and it never comes from missing knowledge: it comes from refusing to be wrong once. Moving it the right way, to protect a gain already made, is an entirely different decision, taken according to a rule written in advance rather than according to how you feel at that moment.

⚠️ A stop that is not in the platform is not a stop. A "mental" stop assumes you will be in front of your screen, available and clear-headed, at the exact moment you will most want to look somewhere else. That assumption is false roughly every time it matters.

R, the unit that makes two trades comparable

Counting in money stops you learning anything. A two hundred pound gain is excellent on a small account, trivial on a large one, and it says nothing about the risk it took to get there. Kept in money, two records compare across nothing at all: not two accounts, not two instruments, not even two months of the same account if the capital moved in between.

R settles this in one move. One R is what you risked on that trade, no more and no less. Take your stop and you lose one R. Exit at three times your stop distance and you bank three R. This is not one more accounting unit, it is a change of rule: instead of measuring the outcome in currency, you measure it in multiples of your own opening decision.

The effect on a record is immediate. An index trade where you risked forty points and a currency trade where you risked twenty-four pips become two rows of the same column. Twenty trades then read as a single addition, and that addition does not change if you double your capital next month. It is the only known way to tell whether your method is improving, independently of the size you put behind it.

⚠️ R is frozen at entry. One R is the INITIAL risk, the one at the moment the order left, never the risk that remains after you have pulled your stop to break even. Otherwise the denominator moves mid-trade and nothing is comparable any more, which is precisely the problem R was brought in to solve.

One last use, often skipped: record the trades you did not take, and the R they would have returned. That shadow record costs thirty seconds a day and answers a question the real record will never ask, namely whether your problem is choosing badly or failing to execute what you had chosen correctly.

Win rate says nothing on its own, expectancy says everything

Win rate is the number everybody quotes, and taken alone it is the least informative of the lot. It answers the question "how often am I right", when the only question that decides the fate of your account is "how much do I make when I am right, and how much do I lose when I am wrong".

The number that answers that question is called expectancy, and its formula fits on one line: expectancy = (win rate × average win) − (loss rate × average loss), all of it expressed in R. The result is what your system returns, on average, per trade taken. A system is only viable if that number is positive, and it never becomes positive because you hoped hard enough.

Two records beat any speech. The first trader wins seven times out of ten, banks half an R per win and hands back two R per loss: his expectancy is 0.7 × 0.5 − 0.3 × 2 = −0.25 R per trade. Over a hundred trades he was right seventy times and he lost twenty-five R. The second wins four times out of ten, banks three R per win and hands back one R per loss: 0.4 × 3 − 0.6 × 1 = +0.60 R per trade, which is sixty R over a hundred trades. The first is right almost twice as often as the second.

That formula also tells you where to look when a system returns nothing, because it offers exactly three levers. You can win more often, win more when you win, or lose less when you lose. Anything that moves none of those three numbers is decoration, and that includes the large majority of indicators added to a chart on the evening of a bad week.

⚠️ Costs enter this formula directly and without negotiation. The gap between the two prices, the commission and the financing charge come off every trade, winners and losers alike. If your edge is worth 0.2 R per trade and friction costs 0.1 R, half your method goes to your broker, and that calculation belongs before opening an account rather than after a hundred trades.

Drawdown, and the asymmetry that makes it dangerous

Drawdown is the fall of your account from the highest point it reached. It is the number nobody displays next to their results, and the only one that describes what you will actually have to sit through in order to get them.

Its mechanics are arithmetic and merciless. The gain required to get back to level equals the loss divided by what is left: losing 10% requires making 11 back, losing 20% requires 25, losing 33% requires 50, and losing half your capital requires doubling what remains of it. The curve does not rise, it runs away, and that convexity is what makes cautious sizing a calculation rather than a temperament.

There are two drawdowns, and confusing them is expensive. Drawdown in R is a property of your method: it is the longest descent of your curve measured in risk units, and it does not depend on what you put on the table. Drawdown as a percentage of the account is what you feel, and it comes from applying your risk per trade to that descent. A method that falls fifteen R costs roughly 15% of the account at 1% risk, where plain multiplication is close enough. At 5% it is not: multiplication would announce 75%, while the account in fact stops a little below half, because every loss is taken out of what is left rather than out of the starting capital. The error always leans the same way, the product rule overstates, and it stays usable while risk per trade stays small.

The real danger of a drawdown is not arithmetic though, it is human. A trader down 30% is not only losing money: he is losing faith in the rule that took him there, changes method at the worst possible moment, and turns an ordinary descent into a permanent change of trajectory. The depth your risk plan allows must therefore be chosen against what you can endure without altering anything, never against what the spreadsheet tolerates.

A losing streak is not a breakdown, it is arithmetic

A run of losses does not announce that your method broke overnight. It announces that you drew the wrong side several times in a row, which is what independent draws do, and they do it a good deal more often than intuition suggests.

The order of magnitude can be calculated. The longest losing run expected over N trades is roughly the logarithm of N divided by the logarithm of one over the probability of losing. Over a hundred trades at one winner in two, the longest run sits around seven. At four winners in ten, around nine. Over five hundred trades at one winner in two, around nine as well. These are not worst-case scenarios, they are the ordinary value, the one you should expect to meet.

The probability of seeing at least one run of six consecutive losses over a hundred trades at a fifty per cent win rate is in the region of one chance in two. Drop the win rate to forty per cent, changing nothing else, and that same probability climbs past four chances in five. The reason sits in the shape of the calculation: one more loss in the run multiplies its probability by the loss rate, so run lengths are paid for in powers while the win rate itself only moved a few points. Nobody will have lied to you about the method: you will simply have drawn a perfectly unremarkable sequence.

The practical consequence is a calculation, not a prayer. Eight losses in a row at 1% risk leave the account at 92% of where it started, which eight winning trades recover. The same eight losses at 5% leave it at 66%, and you now have to make back half of what remains just to return to the starting point. Risk per trade is not a comfort slider, it is what decides whether a normal run is a bad month or the end of the story.

⛔ The reaction that kills is doubling size to win it back. That idea has a name, the martingale, it has been known since the eighteenth century, and its flaw has been proven for nearly as long: it requires an infinite bankroll to work, and it hands you many very small gains before taking everything back in one go. The infinite bankroll does not exist. The one go does.

Risk of ruin, and why 1% is not timidity

Risk of ruin is the probability of falling to a level you do not come back from, given three things only: your edge, your risk per trade and your capital. The question is old. It appears in the correspondence between Blaise Pascal and Pierre de Fermat in 1656 under the name of the gambler's ruin problem, Christiaan Huygens gives a solution as early as 1657 in his De ratiociniis in ludo aleae, and William Feller treats it in the twentieth century in the form we use today.

In the simplest case, where a win pays exactly what a loss costs, the formula is short: risk of ruin equals the ratio of the probability of losing to the probability of winning, raised to the power of the number of risk units your capital contains. An account risking 1% per trade contains a hundred units, an account risking 10% contains only ten, and it is that exponent which does all the work.

Put a modest edge into the formula, fifty-five per cent of winners, and watch what risk per trade does. At 10% risk, ruin arrives in thirteen cases out of a hundred. At 5%, in fewer than two out of a hundred. At 2%, in four cases out of a hundred thousand. At 1%, the number stops having any practical meaning. This is not a slope, it is a cliff, and it explains why a figure that looks excessively cautious is in fact the only one that makes a method indifferent to bad luck.

Watch what the formula assumes, because the confusion is a common one: it describes a constant stake, the kind that cuts your capital into a whole number of identical units, and it is that stake which makes zero reachable in the first place. Risk one per cent of what you hold today instead and the stake shrinks as the account falls, so strict zero never arrives and ruin has to be redefined: it becomes the level you do not come back from, either because you can no longer trade your method at that size or because you no longer have the stomach to follow it. Risking a fixed amount, or worse, a fixed number of lots, does the opposite: the fraction of capital committed grows while the account melts, every loss weighs more than the one before, and that is the mechanism the formula describes faithfully.

The theory has a name, the Kelly criterion, published by John Kelly at Bell Labs in 1956, which gives the fraction of capital that maximises growth over the long run. Two things to take from it. The first is that a fraction exists beyond which risking more earns less, which is not intuitive at all. The second is that this optimal fraction produces falls almost nobody can sit through, which is why practitioners use only a part of it, often a quarter or a half. ⚠️ That fraction is computed on an estimated edge, never on a known one. Thirty trades do not measure an expectancy, they give it a wide bracket, and an overestimated edge pushes the fraction past the point where risking more earns less, which is exactly where the formula stops protecting anything. Take the edge your record measures, keep half of it, and size on that figure rather than on the one you like.

Correlation, or three positions that are really one

Three positions open at the same time, each risking one per cent, give the reassuring impression of spread risk. That impression is false the moment all three depend on the same thing. What you have opened is then not three bets at one per cent, it is a single bet at three per cent, and it will settle all at once.

The commonest cases are easy to recognise once you look for them, provided you look at the direction and not only at the name. Two currency pairs sharing a currency only tell the same story if you are on the same side of it: buying a pair where the euro is quoted first and selling a pair where it is quoted second is buying euro twice over, whereas buying both means buying it on one leg and selling it on the other, which cancels part of the risk instead of doubling it. The method that does not fool you is to write down, for every position, the currency bought and the currency sold, then add up currency by currency. Three shares in the same sector rise and fall with their sector long before they move for reasons of their own. An index and a share that carries real weight inside that index are, from the point of view of your risk, very nearly the same line.

This can be measured. The correlation coefficient, formalised by Karl Pearson in 1896 on the back of Francis Galton's work, runs from −1 to +1: one means the two series rise and fall together, minus one that they always do the opposite. Zero deserves a caveat, because it is where people reassure themselves too quickly: it says no linear relationship links the two series, which is not the same thing as independence. Two instruments can sit happily at zero in normal weather and dive together the day one shared factor hits them both. The coefficient is computed on returns, not on prices, and over a recent window. The simplest working rule is to treat as a single position any group whose correlations exceed 0.7 in absolute value.

⚠️ A correlation is not a constant. It rises when the market panics, because in a panic everyone sells everything at once and the only thing that matters becomes the need for cash. The diversification you measured in calm conditions therefore disappears on exactly the day you were counting on it, and two positions that hedged each other for six months can perfectly well lose together in an hour.

The defence is a ceiling rule, written before the session rather than during it. One per cent per trade, three per cent of open risk at any instant at most, and a correlated group that counts as one single line inside that ceiling. The rule costs nothing on ordinary days, and on the others it is very nearly the only thing standing between you and real damage.

Practise: rebuild your own record in R

None of the above becomes real until you have applied it to your own trades. Here is the exercise. It takes an hour the first time and twenty minutes afterwards, and it runs on your last thirty trades, taken live, on a demo or read back off the chart, it does not matter which as long as they are yours.

For each trade, write three columns: the risk in money at the moment of entry, the outcome in money, and the quotient of the two, which is your outcome in R. That third column is the whole exercise. The first surprise usually arrives right here, when you discover that several losses are worth two or three R when they were supposed to be worth exactly one.

Then compute the four numbers from this lesson: your win rate, your average win in R, your average loss in R, and the expectancy that follows, namely (win rate × average win) − (loss rate × average loss). Write the result on a single line at the top of the sheet, because it is the only figure you will look at in the months that follow.

Finally, find your longest losing run and your deepest descent, both read off the R column. Compare the run against the benchmark in the sixth section: if yours is shorter than the statistics predict, it is not that your method is better, it is that your sample is too small. Then multiply the descent by your risk per trade, and you have what that same method would have cost you as a percentage of the account.

What most beginners discover on running this exercise a single time is always the same thing: expectancy is negative because of three or four trades where R got away, not because of the choice of entries. The fix therefore lives in no indicator, no timeframe and no extra setup. It lives in the risk column, the one filled in before entering. So redo the sheet every month with the month's trades. After six months you will hold a series of measured expectancies rather than an opinion about your level, and you will be able to say, with numbers behind you, whether what you change actually improves something or merely moves the problem elsewhere.