At 00:05 UTC on a Tuesday, our cross-sectional crypto strategy ranked 40 perpetual contracts by a short-term trend score and moved the top ten into equal-weight positions. The backtest showed a healthy spread between winners and losers. Then we asked it to trade $250,000 of capital, and the attractive curve started to come apart.
The trouble wasn’t hidden in the signal. It was in the rebalance: ten buys, ten sells, and a few contracts whose displayed depth looked generous until the orders arrived together. The strategy had treated a portfolio change as a price update. A venue treats it as a queue of actual orders.
We replayed that Tuesday in sequence. First the old positions needed to shrink. Then the new ones had to be opened. The target weights were simple; the path between them wasn’t.
Eight basis points per side came from a single average-cost estimate. It covered the venue fee and a small slippage allowance. That gave every order the same bill, whether it was a liquid contract at a quiet moment or a thin contract just after a burst of volatility.
We started with the largest order. The $25,000 target was 0.4% of daily volume, a ratio that sounded harmless. But daily volume doesn’t tell you what’s available during the seconds when you need to trade. The top of book was thin; the next few levels widened quickly. Crossing for the full amount would sweep several prices. Posting a limit order would save spread only if it filled, and the strategy had no guarantee of getting the fill before its next decision.
So we split the order into five clips of $5,000. That reduced the immediate footprint, but introduced a clock and a market path: each clip could arrive at a different price, and a fast move could leave the position half-built. We recorded both outcomes. The original backtest had recorded neither.
And this was only one leg. The strategy sold positions that had fallen out of the top ten while buying replacements. Its net exposure stayed close to target, but gross turnover was nearly twice the capital allocated. Charging costs only on the final portfolio change would miss the trades needed to get there.
| Execution choice | What it saved | What it risked |
|---|---|---|
| One market order | Time and uncertainty about completion | More book depth consumed at once |
| Five market clips | Smaller immediate footprint | Price movement between clips |
| Limit order | Potentially less spread paid | Partial fill or no fill at all |
That afternoon, we replaced the flat slippage allowance with a simple participation model. Each order’s expected impact rose with its size relative to recent traded volume, and we capped simulated participation so a strategy couldn’t consume an implausible share of the market. It was still an approximation: aggregated volume is not a live queue, and a bar doesn’t reveal every price along the way. But it forced the backtest to admit that the same signal has different execution costs at different sizes.
The revised run also charged taker fees where the order crossed the spread and maker fees only when the simulation actually assumed a resting order filled. Funding remained a separate holding cost; it didn’t belong in the rebalance cost. Keeping those line items apart made the failure legible. The strategy had not suddenly become worse at forecasting. Its turnover was expensive relative to the edge it forecast.
By the end of the week, we had reduced the rebalance frequency and excluded contracts whose available depth made the target weight unrealistic. The backtest’s return fell. Its execution assumptions became easier to explain, which was a more useful result than the bigger number.
When you review a portfolio strategy, follow one rebalance all the way down: from target weights to individual orders, from orders to plausible fills, and from fills to fees and impact. If that trip is missing, the equity curve is describing a portfolio that never had to trade.
← All posts
