You backtested $2,000 a clip and now you want $50,000: a letter about capacity

You sent me the tearsheet and one question: does this survive at size. Five-minute mean reversion, roughly thirty mid-cap USDT perps, every backtest trade sized at a flat $2,000 notional, out-of-sample Sharpe just under 2, average gross edge per round trip around 22 basis points before costs. You've been funding it with pocket money and now you want to put real capital behind it — call it $50,000 a clip.

The honest answer is that capacity isn't a number somebody hands you. It's a curve, and your job this week is to draw it. But you can get most of the way there with a spreadsheet, and I'd rather you did that before you find out the expensive way.

Start with the ratio you already have the data for

Before any impact model, compute participation. Take the quote volume column you already have in your klines — index 7, the USDT turnover of the bar — and ask what fraction of a typical entry bar your order would be.

Here's the shape of it across a universe like yours, using median 5-minute quote volume:

At $2,000 you were a rounding error everywhere, which is why your backtest's fill assumptions never got tested. At $50,000 you are, in the bottom half of that list, the bar. Eighty-three percent of a five-minute bar is not an execution, it's an event. Whatever your fill model says about crossing the spread, it is describing a world you have now left.

Rule of thumb I use for a first pass: above 10% of the bar's volume, stop trusting anything your backtest tells you about that fill. Between 1% and 10%, model impact carefully. Below 1%, the spread and fees dominate and you can move on.

Calibrate the square-root law, roughly, on your own symbols

The workhorse model is the square-root impact law: cost in return terms scales with volatility times the square root of the fraction of daily volume you consume.

impact ≈ Y × σ_daily × √(Q / V_daily)

Y is a constant near 1 for most estimates, σ_daily is the symbol's daily return volatility, Q is your order notional and V_daily the day's notional volume. Take the ARB-shaped name from the table. Median 5-minute turnover of $220,000 is about $63M a day. Daily vol, say 4%.

Run the round trips. At $2,000: 4.5 bps of impact plus 10 bps of taker fees on a 5-and-5 schedule, so about 14.5 bps of cost against 22 bps of gross edge. Net 7.5 bps, which is roughly what your tearsheet shows. Good, the model and your backtest agree at the size you actually tested. At $50,000: 22.6 bps of impact plus the same 10 bps of fees. Cost 32.6 against edge 22. You are paying ten basis points a trade for the privilege.

Set net to zero and solve for Q and you get about $14,000 on that symbol. Not 25x. Seven times, and the last third of that range is a strategy with no margin for error in it. The square root is doing exactly what square roots do: 25 times the size costs you 5 times as much per unit, so the loss compounds against you as fast as the notional grows.

Use the volume you actually trade into, not the median

Here's the refinement that matters more than the choice of Y. You just computed participation against median volume. You do not enter on median bars. Go back and compute V conditional on your signal firing.

Five-minute reversion usually triggers after a violent move, which means your entry bars carry two or three times the median turnover. That's a genuine gift and it can double your real capacity relative to the naive number. Momentum breakouts get the same gift. Anything that enters on a schedule — a fixed rebalance, a time-of-day filter — does not, and anything that enters into quiet drift is trading a thinner book than the median implies.

So measure it. One groupby over your entry timestamps gives you the distribution, and I'd look at the 20th percentile rather than the mean, because the bad fills are what set your risk. If the 20th percentile of entry-bar volume is below the median, your naive capacity estimate was optimistic and you should find out now.

The thing that actually kills scaled-up alt books

Thirty mid-cap perps looks like thirty independent bets in a backtest. In the two hours a month that matter, it's one bet. When a cascade runs, everything correlates to BTC, every book thins simultaneously, and your twelve concurrent $50,000 positions are a single $600,000 directional exposure being unwound into the worst liquidity of the quarter.

Your impact model, calibrated on median conditions, has no idea. Neither does your Sharpe. The fix isn't clever: recompute participation using the 5th-percentile volume day, and check whether your maximum concurrent gross would clear at all. If unwinding the book takes ninety minutes of trading at a 10% participation cap and your edge decays in fifteen, you don't own a five-minute reversion strategy anymore. You own a slow directional position with a reversion story attached.

Every capacity conversation eventually becomes a conversation about holding period. Impact is the price of speed. If you can hold longer you can trade bigger, and if you can't, size is the only variable you get to move.

One more constraint that shows up between $2,000 and $50,000 and surprises people: leverage tiers. On most venues the maximum leverage on a mid-cap perp steps down as position notional rises, and for the smaller names the top bracket often ends well below where you're heading. The margin your backtest assumed at $2,000 is not the margin you get at $50,000. Check the risk-limit table for every symbol in your universe before you assume the capital requirement scales linearly.

What I'd actually do with your week

1. Re-run the same backtest at six sizes — $2k, $5k, $10k, $25k, $50k, $100k — with an impact model wired in that's a function of order size over that bar's quote volume. Not a flat slippage constant. A flat constant is what let you believe in $50,000 in the first place.
2. Plot net Sharpe against size. You'll get a curve that's flat, then rolls over, then falls off. Your capacity is wherever that curve crosses the Sharpe you're willing to run, minus a healthy discount, because every impact model is optimistic in the tails.
3. Split the universe by liquidity tier and size per symbol rather than flat. Flat sizing across thirty names means the tail names are silently subsidising the majors, and they're the ones eating 80% of a bar.
4. Paper-trade at the size the curve gives you, and log realized slippage against the modelled number on every fill. That comparison is the only real calibration you'll ever get for Y. Mine has come back closer to 1.4 than 1.0 on thin alt books.

My guess, sight unseen, is that the answer lands somewhere near $8,000–$12,000 a clip across the liquid two-thirds of your universe, and that you should drop the bottom third entirely rather than accept the fills it will give you. That's still a real multiple of where you are. It just isn't 25x, and finding that out from a spreadsheet costs nothing.

Send me the size curve when you have it. I'm curious where yours rolls over.