I keep running into the same confusion, and it usually shows up when someone shares a backtest with a Sharpe near 3 and a return curve that looks like a rocket launch. My first question is never about the signal. It is about how they sized the trades. Because the same exact entry and exit stream, run through three different sizing rules, will hand you three completely different stories about how good the strategy is. One version looks flat and boring, another looks decent, and the third looks like something you should quit your job over. Same signal. The difference is entirely in the staking math.
This matters because sizing is the easiest place to fool yourself. You can build an honest signal, size it aggressively, and end up with a backtest that says far more about compounding than about whether you actually predicted anything. And most people never separate the two, so they cannot tell you which part of the pretty curve is skill and which part is leverage dressed up as skill.
Three sizing rules, three different backtests
Start with the simplest case. You have a signal that goes long when some condition fires and flat otherwise. Now pick how much to stake.
- Fixed units. Every trade is the same absolute size. One share, one contract, one coin, whatever. Your account can grow ten times over and you are still trading the same tiny slice. This is the most conservative and the most honest reflection of the signal itself, because each trade contributes to the record on roughly equal footing.
- Fixed fractional. Every trade risks the same percentage of current equity. As the account grows, position sizes grow with it. Losses shrink your bet after a bad run and expand it after a good one. This is where compounding sneaks in, and where a mediocre signal can start to look impressive.
- Full compounding. You reinvest everything, so each trade is sized off the entire current balance. On a winning strategy this produces the exponential curve everyone loves. It also magnifies the ordering of your trades in ways that have nothing to do with signal quality.
Run one signal through all three and watch what happens to the headline numbers. Total return diverges the most, obviously, because that is just the compounding curve doing its thing. But the sneaky part is what happens to Sharpe and to drawdown. Fixed fractional and full compounding tend to flatter your risk-adjusted numbers during the equity ramp, because your bets are largest exactly when the strategy has been winning, and they shrink right before it stops working. When the ordering of wins and losses happens to be favorable, this looks like genius. It is mostly luck about sequence.
Why the ordering of trades starts mattering
Under fixed units, the order of your trades barely affects the final result. Ten wins and ten losses net out to roughly the same place no matter how you shuffle them, because every bet is the same size. That property is exactly what you want when you are trying to judge a signal, because it means the record reflects the trades and not the sequence.
The moment you size off equity, sequence becomes load-bearing. A cluster of early wins gives you a bigger bankroll for every trade that follows, so late wins get amplified. A cluster of early losses does the opposite. Two backtests with the identical set of trades, just reordered, can post noticeably different Sharpe ratios and very different max drawdowns under compounding. That is not a signal property. It is an artifact of the staking rule interacting with one particular historical path. Change the path, and the appeal evaporates.
This is also why aggressive compounding backtests are so fragile out of sample. The historical sequence gave you a friendly ordering. Live trading gives you a different ordering, and the magic disappears while the drawdowns stay.
Evaluate the signal first, size it second
The fix is to treat these as two separate decisions and test them separately. First ask whether the signal has any edge at all, using sizing-neutral measures. Then, and only then, ask how to size it.
For the first question, I lean on metrics that do not care about compounding. Average return per trade in the raw asset, hit rate, average win versus average loss, and expectancy per unit of exposure. Look at the distribution of individual trade outcomes, not the equity curve. If the per-trade edge is thin or inconsistent, no sizing rule will save it, and any pretty curve you build on top is borrowed from the compounding math, not the signal.
A rough workflow I actually use:
- Run the strategy in fixed units first. If it does not look appealing here, the edge is weak and everything downstream is decoration.
- Compute per-trade expectancy and the spread of outcomes. Confirm the edge survives once you strip out sequence effects, ideally by reshuffling trade order a few times and checking the numbers stay stable.
- Only after the signal earns its keep, layer on a sizing rule and treat that as its own experiment with its own risk budget.
- When you do compound, report the fixed-unit numbers next to it so a reader can see how much of the appeal came from staking versus from the trades.
The reshuffle test in step two is the cheap one that catches the most self-deception. Take your list of trades, randomize the order a handful of times, and recompute Sharpe and drawdown under your compounding rule each time. If those numbers swing wildly, your backtest is riding a lucky sequence and you should trust the fixed-unit version instead.
Reading someone else's backtest
When a backtest lands in front of me, sizing is the first thing I try to reverse engineer. A few tells. If total return is enormous but the per-trade edge looks ordinary, that gap is compounding, not skill. If the equity curve is smooth and exponential but the drawdown number seems suspiciously mild, ask whether bet size was shrinking into the losing stretches, which flatters drawdown without protecting you against a differently ordered future. If nobody can produce the fixed-unit version of the same strategy, that absence is the answer.
None of this means compounding is bad. It is how accounts actually grow, and a real edge deserves to be sized up. The problem is only that sizing and signal quality get blended into one number and then presented as if the number measures the signal. It does not. It measures both at once, and usually the flashier the curve, the more of it is sizing.
When we built the strategy tester inside Blockcircle, this is roughly why we default to showing per-trade stats alongside the compounded curve rather than leading with the total return. It is less impressive at a glance, and that is the point. You want to know what you actually have before you decide how hard to press it. Test the signal until it either earns your trust or does not, then make sizing a deliberate choice you can defend on its own, with its own drawdown you have looked at with clear eyes.