Technical Analysis · Chapter 12 of 12
From a rule to a system
Entry is the part everyone works on and the part that matters least. Here is where the expectancy actually lives.
A signal is not a strategy. A strategy is a complete specification of what you do in every state, and the components that determine your results are, in rough order of importance, almost the reverse of where people spend their attention.
The five components, in order of how much they matter
| Component | Why it matters | Where people's effort actually goes |
|---|---|---|
| 1 · Position sizing | Determines the variance of your equity curve and whether you survive the drawdown. Two traders with the identical signal and different sizing have unrecognisably different outcomes | Almost none |
| 2 · Exit / risk rule | Where most of the expectancy lives. Cutting losses and letting winners run is the entire distribution of a trend system | A little |
| 3 · Universe and regime filter | Deciding when not to trade is usually worth more than improving entries | Some |
| 4 · Entry | Matters, but far less than the three above | Almost all of it |
| 5 · Parameter values | Should barely matter. If they matter a lot, the strategy is fitted | A great deal, unfortunately |
Sizing, concretely
The volatility-scaled rule from chapter 7 is the default: risk a fixed fraction of capital, and let ATR determine the quantity. Beyond that, three points worth knowing:
- Fixed fractional risk compounds correctly. Risking 1% of current capital means your position shrinks after losses and grows after gains, which is the behaviour you want and is not what a fixed quantity does.
- Correlated positions are one position. Six long positions in the same sector at 1% risk each is a 6% risk position wearing a diversification costume. Cap risk by correlated group, not only per trade — this is the single most common way a well-designed system still blows up.
- The Kelly criterion gives the growth-optimal fraction and nobody trades it. Full Kelly produces drawdowns that no human tolerates and is extremely sensitive to your estimate of the edge — and your estimate is wrong. Half-Kelly or less is the practical convention. Knowing why you would not use full Kelly is a better answer than knowing the formula.
Exits, where the expectancy hides
Most people specify an entry precisely and an exit vaguely. It should be the other way round, because the exit determines the entire shape of your return distribution.
- The initial stop should be at a level that means the idea was wrong, not at an amount you are comfortable losing. Those are different things and conflating them produces stops that get hit by noise. Structure first, then size the position so that structural stop costs an acceptable amount — if it costs too much, the trade is too big, not the stop too tight.
- A fixed target caps your winners, which is fine for a mean-reversion system and fatal for a trend system, where a small number of very large winners carry everything.
- A trailing stop lets winners run at the cost of giving back some of the gain. The trade-off is real and it should be measured rather than assumed: test both on the same signal and look at the whole distribution, not the average.
- A time exit is under-used and often improves results. If the idea has not worked within the horizon you expected, the reason for holding is gone.
- Move to break-even too early and you will convert winners into scratches. This feels prudent and is usually costly. It is a very testable question and the answer is frequently the opposite of the instinct.
The regime filter, which is cheaper than a better signal
Most strategies work in some conditions and lose in others. Identifying the conditions and simply not trading in the bad ones is usually a larger improvement than any amount of entry refinement — and it is more robust, because it has fewer parameters.
Practical filters worth testing: a long-term trend filter (only long above the 200-day); a volatility regime filter (ATR relative to its own history); a liquidity filter (skip anything below a turnover threshold, which also solves half your slippage problem); and an event filter (no new positions immediately before earnings or a scheduled announcement).
What to do when a live system stops working
It will. The question is whether it is noise, decay or a break, and the three need different responses.
Decide in advance what would constitute failure — a drawdown exceeding the worst in your backtest, or a rolling expectancy below zero over a defined number of trades. Write the number down before you are in the drawdown, because in the middle of one you will not be able to think about it clearly, and every discretionary decision at that point will be made by the part of you that wants the pain to stop.
Noise looks like a drawdown within historical range. Do nothing. This is by far the most common case and doing nothing is the hardest correct action in trading.
Decay looks like expectancy drifting toward zero over many trades. The edge is being competed away. Reduce size, and start looking for the next thing.
A break looks like behaviour outside anything in the historical record, usually coinciding with a structural change — a rule change, a new contract specification, a change in market composition. Stop, and re-derive rather than re-fit.
What to actually do next
- Pick one rule from this book and test it properly, all five steps. One honest test teaches more than fifty charts.
- Keep a research log. Every hypothesis, the result, and the date — including the failures. This is the artefact that makes your denominator honest, and it is the single most persuasive thing you can show in a quantitative interview.
- Paper-trade the mechanical version before risking money, and compare the live fills to the backtest's assumed fills. The gap between them is your real cost model, and it is always worse than you assumed.
- Then start small, and scale only when live results match the tested distribution — not when they exceed it, because exceeding it is also evidence that your model is wrong.
Check yourself