The lock window, and what it costs to keep

Go to a racetrack and try to bet ninety seconds after the gates open. You can't. Betting closed at the off. Go to a sports pool and try to enter at half-time on a 3–0 scoreline. You can't. Entries closed at kick-off. Go to a prediction market running a pool and try to buy the obvious answer thirty seconds before the announcement. Most of the time, you can't.

Every venue that runs a parimutuel closes it early. This is so universal that it reads as a fact of nature, like a market having opening hours. It isn't. It is a patch for a specific defect, and the patch is expensive.

The defect

A classic pool pays a flat multiple. If the winning side holds $10,000 out of a $25,000 pool, everyone on that side is paid 2.5× — the person who committed at the open, and the person who committed after the outcome was effectively known.

That second person has a strategy. Wait until the event is nearly decided, identify the side that is now overwhelmingly likely to win, and put money on it. You are buying a near-certainty at pool odds set when it was still uncertain. Your money joins the winning book, which reduces everyone else's multiple, and you collect a return you took no risk to earn.

This is not a theoretical exploit; it is the dominant strategy, and it has been studied experimentally since Plott, Wit & Yang (2003). The paper measures it directly. A sniper stakes 50% of the gross pool on a leader above 85% at t = 0.95T, across 20 seeds × 2,000 markets (§13.1):

A +25% expected-value strategy available to anyone with a clock is not a thing a venue can leave open. So it doesn't.

What the patch costs

Handle. The most obvious cost, and the one an operator feels in the P&L. Every minute of closure is a minute of volume the venue does not take. In sports this is the worst possible window to be shut: interest peaks near the event, and the venue is dark for exactly that peak. In the paper's behavioural study the lock arm loses 5.0% of its total volume relative to the unlocked arm — and that is with an artificial crowd. Real venues close for far longer than the final 5% of a market's life.

A closed market is a dead market. The lock does not just stop the snipe. It stops the hedger with real exposure, the trader whose thesis matured late, the person who only just heard about it. Everyone gets locked out to stop one strategy.

It is a policy, not a property. Someone has to choose the window, defend it, and adjust it per market class. It is a parameter with a business case attached, and it is the venue's fault when it is wrong in either direction.

It does not actually fix dilution. This is the part usually missed. The lock stops the buzzer snipe. It does nothing about ordinary late money. Look at the classic row of the paper's payout-by-decile table, 20 seeds × 2,000 markets:

The classic row is flat, and slightly rising at the end. Being early is worth nothing at all — and the last decile of entrants captures 13.20% ± 0.07 of the losing pool. That transfer is happening inside every locked market, every day, below the level the lock operates at.

What replaces it

The vested rule makes the snipe pointless rather than illegal, and it does so as an identity rather than as a measure.

Here is P4 in full:

A stake accepted after the last opposing inflow is paid exactly its principal; for a trader with belief p < 1, its expected value is −(1−p)s < 0.

The proof is two clauses. Vesting only ever adds increments from opposing inflow, so if there is no opposing inflow after you, you accrue nothing and are paid your stake back. And under the capacity rule there is no unmatched residue to recover — a classic pool's late entrant at least gets 1× less takeout on unmatched money, which is a soft floor; here the negativity is unconditional.

So: 1.000× on the win, 0 on the loss. Any belief short of certainty makes that strictly negative. The strategy has no target, and the venue does not need a rule to stop it.

That is the corollary the paper draws: no lock window. Markets can accept entries right up to resolution, because the reason to close them has been removed rather than policed.

The number that matters is the pair

A single number can be argued with. The paper's approach here is to publish the signed pair on the same triggers: +25.27% under classic rules, −1.05% under this one. Same crowd, same markets, same sniper, same 98.42% base rate on the leader.

That pair — not the lock rule, and not a claim about anyone's intentions — is what deletes the strategy.

There is a residual, and the paper does not round it away. Early winners still lose 0.19% under the snipe. That is not a leak. It is early winners sharing the last 5% of opposing flow with a newcomer who genuinely showed up and genuinely took the other side of it. That is P3 working exactly as specified: your accrued claims are untouched, but a later same-side entrant does share the flow that arrives after them.

What it looks like when the crowd responds

Settling a fixed log under two rule sets proves the arithmetic. It does not tell you whether people change their behaviour. So §13.2 re-runs the experiment with agents deciding for themselves whether to enter, under whichever rule they face.

Share of volume by phase of market life:

Late pool volume collapses from 5.1% to 1.7% without prohibiting anything, and a third more volume migrates into the market's opening phase. The conservative-estimator arm (agents using a deliberately pessimistic growth assumption) pushes further in the same direction — 48.5% in the first third — so the migration is driven by the incentive, not by a particular estimator.

The residual 1.7% is worth naming honestly. It is not zero, because a market's last dollars are not all buzzer snipes; some late entrants still have positive expected value because opposing flow is still arriving behind them.

What it costs to delete it

This is where an honest article has to turn around, because the trade is real and the paper measures it in the direction that hurts.

If rational late money stops entering the pool, the pool's ratio stops updating near the end. It goes stale. And a stale ratio is a worse forecast:

Lower is better. The fair comparison is the middle row against the bottom: the deployed baseline, not the ideal one. The mechanism's late forecast is measurably worse — by 0.024 Brier — and the paper's own summary of the trade is that it "pays forecast quality in the pool's last act to buy dilution protection for everyone before it."

Two things soften that, neither of which is a dismissal.

The information still exists; it just stops showing up in the pool. Adding a stylized dealer — every late informed arrival prints a transaction at its belief, shaded by half a spread — produces a composite forecast with a final-phase Brier of 0.054–0.057, at every spread from 2% to 20%. Better than the locked classic pool by a wide margin. That is an upper bound on what a thin transaction layer could deliver, not a design; the article on the late price covers what is and is not established there.

The lock arm has the same problem, one step earlier. A locked pool's "final" price is simply its price at lock time, frozen. Its 0.150 is a snapshot taken before the last 5% of information existed. The mechanism at least keeps accepting entries; what it loses is the incentive for informed ones.

The one-line version

A lock window is a venue admitting that its settlement rule pays people for showing up late, and closing the door rather than fixing the rule. Fixing the rule removes the door — and moves where the last hour's price comes from, which is a cost, stated at 0.174 against 0.150.

Next: Who funds the first counterparty — the objection at the other end of a market's life.