Article · 15 min read · 2026-05-22
Edge is in the Process
Hold every strategy and you lose. Hold the selected few and the same losing pool turns profitable. Selection, not the strategies, is the edge.
The whole pool loses. The selected few do not.
Here is the result first, because it is the point. I evaluated 533,638 distinct strategies across 12 crypto asset/timeframe universes, every number below is net of costs (0.05% taker fee, 0.03% slippage per fill, 0.01% funding per leg). Equal-weight all of them and the portfolio's median profit factor is 0.48: a reliable, grinding loser. Rank that same pool by past profit factor and hold only the top handful, and the median climbs to 0.85–0.96, and on 8 of the 12 markets it crosses 1.0 outright, turning profitable. Same strategies, same data, same costs. The only thing that changed is the rule that decides which ones to hold.
That gap, from a losing pool to a profitable portfolio, is the entire article, and none of it needs a black box. The selection rule is arithmetic on realised profit factors, walk-forward and reproducible. What I will also show is that the rule used here is deliberately crude, and that its limitations, not the concept, are what stand between this and a robustly profitable book.
Why you can't just pick the winner
The reason you cannot skip selection and simply buy "the good strategy" is that, at the level of the individual strategy, there is almost nothing to pick. Take any strategy's profit factor in one 4-week window and ask how it does the next: the rank correlation, pooled across all 44 million strategy-windows, is about 0.02. A single strategy is a lottery, its median next-window profit factor is 0.82, but the distribution is enormous, and which ticket wins is close to unknowable in advance.
If last window's profit factor predicted next window's, the green conditional-median line would ride the diagonal. It is flat, near 1, regardless of in-sample performance. This is the trap retail strategy-hunting falls into: the backtest that looks brilliant is a coin that landed heads, and the coin has no memory.
Across 44 million strategy-windows, last window's profit factor predicts next window's with a rank correlation of 0.02. You cannot pick the winner. You can only select on the population.
The setup
The design is deliberately exhaustive so nothing hides in a clever parameter. Each market has 30,000–50,000 parameterised strategies (ATR, EMA, MACD, PPO, RSI, SMA, STOCHK families and their stop-loss / risk-reward variants). I cut each market's timeline into consecutive walk-forward windows of 28 trading days and compute, per strategy per window, a profit factor, up-day PnL over absolute down-day PnL. Every selection decision for a window uses only realised PnL from strictly earlier windows; nothing looks across the boundary it is measured at. Throughout, I report the median profit factor across windows rather than the mean: profit factor is a right-skewed ratio, and the median is the honest description of the typical window.
Four selection rules
At each forward window, from the same eligible pool, I build an equal-weight portfolio four ways. Only the membership rule changes.
- None, hold every strategy active in the prior window. No selection: the raw pool.
- Random N, N strategies drawn at random.
- Top-N by last window's PF (single-window selection), rank by the prior window's profit factor; keep the top N.
- Top-N stable (multi-window selection), keep only strategies with profit factor > 1 in all four trailing windows, rank those by aggregate profit factor, keep the top N, and walk forward keeping a four-window memory.
How much better is selection?
Line the rules up on a single axis, median profit factor, and the ladder is stark.
| What you hold | Median PF | Gap to break-even closed |
|---|---|---|
| Raw pool (hold everything) | 0.48 | n/a |
| Random 25 | 0.61 | +24% |
| A single strategy | 0.82 | +65% |
| Selected top-10 (stable) | 0.85 | +72% |
| Tightest selection | 0.96 | +92% |
The raw pool sits at 0.48 because the average strategy loses after costs, and equal-weighting thousands of them just averages the losses into a steady bleed. Selecting the top ten stable strategies lifts the median to 0.85, closing 72% of the distance from that losing pool to break-even, and the tightest selection reaches 0.96, 92% of the way. Random selection does the opposite of skill: it decays as the basket grows (0.83 at N=1 down to 0.54 at N=100), because a bigger random sample just regresses to the losing pool.
Selecting the top ten strategies closes 72% of the gap between a losing pool and break-even. The strategies didn't change. The question you asked of them did.
Filter harder, and it crosses into profit
If selection is doing the work, filtering harder should do more of it, and it does, monotonically.
Pooled across all twelve markets the curve approaches 1.0 but, with this crude signal, stops just short. Look market by market, though, and the story is sharper: under tight selection the median portfolio is outright profitable on 8 of the 12 markets, XRP at 1.22, ZEC and DOGE near 1.09, ETH, BCH and BTC just above 1.07, TRX and SOL over 1.0.
This is the answer to the obvious next question: could you make the whole thing profitable if you could filter better? The evidence says yes, and you can see the mechanism. The ranking signal used here, last window's profit factor, carries a rank correlation of just 0.02 with the future. It is almost pure noise, and it still crosses into profit on two-thirds of the markets, because even a weak signal, applied to the population, concentrates the pool's hidden positive mass. The four markets that stay below 1.0, and the goal of carrying a larger, less fragile portfolio across rather than a tightly-concentrated handful, are not limits of the idea. They are limits of the signal. A sharper rule than "rank by past profit factor" moves the whole curve up.
A selection signal with a 0.02 correlation to the future already turns the pool profitable on eight of twelve markets. The ceiling is not the concept. It is how well you can rank.
Selection buys consistency, too
There is a subtler reason a selected portfolio beats raw strategies, and it survives even where the medians are close. A single strategy is not just a coin flip in expectation, it is wildly dispersed. A selected portfolio of the same median quality is concentrated: it does roughly the same thing every window instead of swinging between windfall and ruin.
That tightness is what makes the next step, diversifying across markets, pay off.
Stacking pairs: a second layer of selection
Treat each market's selected portfolio as one instrument and combine across markets. The same theme, selection, returns as which pairs you combine. To make the contrast concrete the simulator includes a high-correlation anchor, one market's selected strategies split into two baskets, which are naturally correlated, against combinations of genuinely different markets.
The relationship is about as clean as empirical finance gets: across all 78 combinations, the volatility removed by an equal-weight combination tracks the correlation of its pieces with a Pearson r of −0.97.
Because the selected portfolios are nearly uncorrelated across markets, stacking all twelve cuts portfolio volatility by 68%, saturating around eight pairs. Selection got each market to the edge of profitability; combining low-correlation selected markets is what makes the result something you could actually hold.
What this claims, and what it doesn't
The practical reading:
- A library of strategies with no selection rule is a liability, not an asset. Held together, the 533,638 strategies lose money; the value appears only when a rule decides which to hold.
- The unit of analysis is the population, not the strategy. Individual persistence is 0.02; the selection lift over the pool is 72% of the way to break-even and crosses it on most markets.
- Better selection, not better strategies, is the growth axis. The concept is proven on two-thirds of markets with an almost-noise signal. Sharpening the rank is what carries the rest, and a larger portfolio, into profit.
Reproduce it yourself
Nothing here needs more than a per-strategy daily PnL table. Cut each market into fixed walk-forward windows; compute each strategy's per-window profit factor; at each window rank the prior window(s), hold the equal-weight top-N, and measure the next window's median. Compare against holding everything and against random draws. The ladder in Figure 2 and the per-market crossings in Figure 4 will reappear. The edge was never a secret strategy. It was the selection process, and it is yours to sharpen.