LZLZL/Prediction markets/Engineering
FREEBUILD IT G · EngineeringSimulation

Walking the book
quotes are not executable size

2026-08-21 · why a backtest priced at the mid is unreproducible

A quote tells you the price of the first few shares, not of your order. Want more and you fill level by level, at a worse average. Backtests that ignore this manufacture opportunities that never existed — and on this venue the books are thin enough that it happens constantly.

1How thin, measured

Full-level replay across more than 1,600 short-tenor windows:

PercentileTop-of-book size
medianabout $94
25th$37
10th$13

Heavily right-skewed: the median looks workable and one window in ten has only $13 at the best price. An order of any size is walking the book on those.

2What a single order actually costs

Taking a realistic order size against those real books:

MeasureResult
Windows filled with zero slippage45%
Average slippage0.56pp
Windows that could not be filled0%

So it is fillable — but 55% of the time it costs you something, and that something averages more than half a percentage point. Against a cost line already around 2pp, this is not a rounding error.

3★ Where capacity actually ends

BasisOrder size at which net return reaches zero
Counting book-walking onlyabout $200
Also counting price drift during executionabout $90
⚠ Why "the backtest was great, so I will size up ten times" nearly always fails

Because an edge has a capacity, and on short-tenor boards that capacity is often only a few hundred dollars. Size up tenfold and the edge may already be gone or negative — the strategy did not stop working, you overran it.

The right question is not "how much does this make" but "at what size does it still make it". Those are different questions, and a backtest answers only the first by default.

4⚠ Splitting the order does not help

The instinctive fix is to break a large order into pieces. Measured, it is strictly worse, for two separate reasons:

Split howResult
Simultaneouslyyou consume the same levels, so the volume-weighted price is unchanged
Spread over timeyou pay price drift — measured at about 0.79pp per 30 seconds, against roughly 0.46pp of book-walking for taking it in one go
take it in one order, and take it fast drift costs more than depth does

5⚠ And sizing to depth is a trap

⚠ "Scale down when the book is thin" is hidden selection

It sounds prudent. But depth correlates with price, and price correlates with the kind of window you are in — so adjusting size by depth means you are silently reweighting your bets by a market characteristic.

That is a strategy change disguised as a risk control, and it belongs to the same family as capping your fill price: a rule that quietly selects which windows you participate in, without appearing in your strategy description.

If you want depth-dependent sizing, declare it as part of the strategy and test it as such — do not smuggle it in as execution hygiene.

6What a backtest must do

MustBecause
Fill against real levelsa flat price produces a result you cannot reproduce
Include execution driftit halves capacity: $200 becomes $90
Report no-fills as a ratedropping them silently is selection on the outcome
State the bias directionconservative or optimistic — see replay bias
Test at the size you will tradecapacity is a property of the edge, not a detail

7One line to keep

A quote is the price of the first few shares, not of your order. On these books one window in ten has $13 at the top, and the size at which an edge dies is about $90 once drift is counted — which is why sizing up is where good backtests usually go to die.

EvidenceCheck it yourself

Depth full-level replay across more than 1,600 windows: top-of-book median about $94, 25th percentile $37, 10th percentile $13.
Single-order cost 45% of windows filled with zero slippage, average 0.56pp, 0% unfillable.
Capacity net return reaches zero at about $200 counting book-walking alone, and about $90 once execution drift is included.
Splitting simultaneous splits consume the same levels leaving the weighted price unchanged; splitting over time costs about 0.79pp per 30 seconds of drift against about 0.46pp of walking in one go.
Checked 2026-07 to 2026-08.
Not included my strategy parameters and any per-window selection rules.

NextWhere to go

F · METHOD
The real ceiling on each path
F · METHOD
Replays must state their bias
B · REBATES
Queue position and conservative assumptions
C · ARBITRAGE
Where paper spreads actually live
This is an educational and research record. It is not investment advice, promises no returns, and offers no personalised trading recommendations. Rules and API behaviour are per the official documentation; this page states when it was checked and both can change without notice. Prediction markets are restricted or unavailable in some jurisdictions — confirm your own before taking part.