One Hundred Strategies That Work
Published 2026-07-25 · By Drew Shelem
A companion to The Best of N Is Not a Result. It covers a case the main article left aside: you do not have one selected strategy, you have a portfolio. Every number is reproducible — script attached.
If you read nothing else: averaging across a portfolio removes the spread and does not remove the bias. A hundred strategies, each slightly inflated, produce a portfolio inflated by exactly the same amount — while looking far more convincing than any one of them alone.
The situation
The previous article covered the case where you sweep many variants and keep the best. Readers point out, fairly, that this is not how they work.
The real process looks different. You come up with an idea. You test it. You nudge the parameters a little. If it works you keep it; if not you move on to the next idea. Six months later you have a hundred strategies, and all hundred work in backtest. You did not pick the best of a thousand — you built a hundred of them, one at a time.
Then you assemble a portfolio from them. The equity curve comes out smooth: a hundred sources of return, drawdowns that do not coincide, everything averaged out. It looks far more dependable than any individual strategy.
Let us go through what is true here, what is not, and what the variations are.
Good news: your N is smaller than you thought
If you genuinely did not choose between ideas, the correction from the main article does not fully apply to you.
There was no competition between strategies, so there should be no penalty for one. Your N is the number of configurations swept inside a single strategy, not the total number of ideas.
pick the best of 100 ideas → inflation of 2.53 sigma
keep all 100, sweeping 10 configurations
inside each one → inflation of 1.57 sigma
The difference is substantial, and it is honestly earned: you were not running a contest, you were building one at a time.
That is where the good news ends.
The bias does not average out
The natural thought: errors across a hundred strategies should partly cancel. Lucky here, unlucky there, and it evens out.
It does not. I checked by simulation: a hundred strategies with a true edge of zero, ten configurations swept inside each, best one kept.
| Configurations inside each strategy | Portfolio in backtest | Portfolio live |
|---|---|---|
| 1 | 0.00 | 0.00 |
| 5 | +1.16 | 0.00 |
| 10 | +1.54 | 0.00 |
| 20 | +1.87 | 0.01 |
Every strategy is inflated by 1.54. The portfolio is inflated by the same 1.54.
The reason is simple and worth spelling out. Errors cancel when they point in different directions. Here they all point one way: inside each strategy you kept the configuration that showed the best result. Not a random one. The best one. So every strategy is biased upward, and the average of a hundred upward biases is an upward bias.
Averaging removes what is random. Selection is not random.
And here is the trap
Diversification does exactly what you expect it to — to the spread:
| Strategies in portfolio | Bias | Spread of result |
|---|---|---|
| 1 | +1.54 | ±1.00 |
| 10 | +1.54 | ±0.32 |
| 100 | +1.54 | ±0.10 |
The spread fell tenfold. The bias did not move.
Put the two columns together and the picture is unpleasant. A portfolio of a hundred strategies shows a consistently inflated result. Smooth curve, even months, no outliers — and a systematic error the same size as a single strategy’s.
One noisy strategy at least jumps around honestly, and that provokes healthy doubt. A hundred averaged together produce a flat line that invites belief precisely because it is flat.
The smoothness of the curve is evidence of diversification, not of being right. Those two things are easy to confuse, because in textbooks they usually arrive together.
Variation one: you discarded some ideas
Now check yourself honestly. How many ideas did you test and throw away?
If you tested exactly a hundred and kept all hundred, there was no selection between ideas, and everything above applies.
If you tested a hundred and thirty and kept the hundred that “worked,” there was selection. Mild selection — the works/does-not-work filter cut thirty — but selection. For a claim about the portfolio, your N is a hundred and thirty, not ten.
The difference between these cases is not how strict the filter was, but whether one existed at all. “I keep it if the backtest is positive” is a threshold too, and it selects too.
Variation two, the most common: the stopping rule
This one is subtler and leaves no trace in any counter.
You took an idea. Ran it with default parameters — nothing. Nudged the threshold — better. Once more — it worked. You stopped and filed the strategy under “working.”
Three runs. So N = 3?
Ask yourself: if it had not worked on the third, would you have stopped? Almost certainly not. You would have tried a fourth, a fifth, a tenth. Which means you were prepared to go further and stopped only because it worked.
Here is what that does to the result:
| Prepared to go up to | Expected result at zero edge |
|---|---|
| 3 runs | +0.85 |
| 5 runs | +1.16 |
| 10 runs | +1.50 |
| 20 runs | +1.77 |
The inflation is set by how many runs you were willing to make, not by how many you made. Stopping on a lucky result is selection too; it just looks like saving time.
Practical consequence: decide how many configurations you will test before you start. Then N is known and the stopping rule stops drifting.
Variation three: your hundred strategies are not a hundred bets
Last one, and this is about diversification rather than selection.
Your strategies differ by construction — different indicators, different logic. But they trade the same markets. If three strategies buy the same contract at the same moment for three different reasons, that is not three bets. It is one bet at triple size.
How much this changes things:
| Correlation between strategies | Spread of a 100-strategy portfolio | Effectively independent bets |
|---|---|---|
| 0.0 | 0.100 | ~100 |
| 0.1 | 0.325 | ~9 |
| 0.3 | 0.562 | ~3 |
| 0.6 | 0.765 | ~2 |
Notice how fast this happens. A correlation of 0.1 — very weak, essentially invisible to the eye — is enough for a hundred strategies to behave like nine. At 0.3, like three.
Now combine that with the previous section. You attribute the smoothness of your curve to diversification across a hundred sources, when in fact there are three. So the smoothness deceives twice over: it does not cure the bias, and there is less diversification behind it than it appears.
This is directly checkable: compute the correlation matrix of your strategies’ daily returns. The average correlation is your answer.
What to do about all this
Count three numbers instead of one.
- How many configurations were swept inside a typical strategy — that is your N for a claim about an individual strategy.
- How many ideas were tested and discarded — that is the addition to N for a claim about the portfolio.
- The average correlation between strategies — that is how much smaller your portfolio is than it looks.
Examine your stopping rule. If you stopped sweeping when it worked, N is not the number of runs but the extent of your willingness. The honest estimate is the pessimistic one.
Do not trust smoothness. A flat portfolio curve means the strategies’ drawdowns did not coincide in the past. It does not mean the strategies have an edge, and it does not mean the drawdowns will not coincide in future.
And above all, the same point as the main article. The only clean test of a portfolio is data that did not exist at the moment of selection. Not half of the history you already have — new history. For a portfolio this matters more than for a single strategy: the bias is the same size and it looks more convincing.
What this discussion does not cover
Position sizing and strategy weights. Equal weights were assumed throughout. Weighting by backtest results is another round of selection, and it adds its own bias.
Coinciding drawdowns. The correlation above is an average, and on prediction markets it typically rises under stress: at the moment the market behaves unusually, strategies tend to be wrong together. Diversification weakens exactly when it is needed.
Selection on stability. If you kept strategies not for their returns but for how “stable” the result looked, that is selection too — and what gets biased is your estimate of risk rather than of return.
Not financial advice. Past results do not predict future results. Trading prediction markets carries risk up to and including total loss of capital. Verify that trading on these venues is legal in your jurisdiction.