FINDINGS.md

# Polymarket up/down crypto markets — what actually has edge (honest findings)

*Companion to the `polymarket-updown-microstructure` dataset. 2026-06-25.*

This is a negative-results report, written to the standard that matters on an
efficient venue: **every edge claim is graded against the price you could actually
transact at**, not win-rate. On a market where the ask already encodes P(win),
a high win-rate is not edge — `edge_real = realized_WR − avg_price_paid` is.

## TL;DR

Across an exhaustive search of taker signals, a full physics-regional sweep of the
maker side, an active cancel-race simulation, and replication against two independent
external corpora, **no signal-based strategy clears the realizable bar** — taker or
maker, passive or active. The informed flow leads the settlement oracle, so even a
*perfect, zero-latency* oracle-based maker cancel does not help. What edge exists on
this venue belongs to actors racing inside the inter-trade gap on non-public signals;
it is not reachable from price/oracle data, at any latency we measured.

## 1. Method

- **edge_real** = `(won ? 1 : 0) − price_paid`, per fill, bootstrap-CI'd.
- **Realizable (fillable) lens**: a backtest "fill" counts only if a *real* trade
  print corroborates it (tape) or the quote survives our measured latency
  (persistence). Paper-fills at the displayed quote overstate edge ~10× via
  fill self-selection (you fill the losers, get rejected on the winners).
- **Ground-truth resolution** from the settlement oracle, never from spot capture.
- Calibration anchor: a strategy with a known live `edge_real ≈ −0.004` reproduces
  to that value under the lens before any verdict is trusted.

## 2. Taker side — closed

An exhaustive grid (5,161 configurations: entry-timing × price-band × move ×
fade/decisive, across BTC/ETH/SOL/XRP, 5m/15m) graded on the realizable lens, then
an adversarial verification stage (pooled-CI + persistence-agreement + date-holdout
OOS):

- The overfit cells evaporate under pooling (a family at +0.12 on n≈50 per cell is
  **−0.017** pooled to n=1752).
- Most apparent survivors die on persistence-vs-tape disagreement (tape says +,
  the resting book says −: the fills are phantom/adverse).
- Exactly one configuration survived the full gauntlet, and it was carried entirely
  by a single coin at n=63 — a pocket, not an edge — and fires ~1×/day (economically
  negligible even if real).

Live confirmation: every directional taker we shipped died live (momentum, entropy,
velrider, flip), each time because the backtest measured `corr(signal, observed
price)` while live fills self-select into the losing side.

## 3. Maker side — no physics region

The passive (post-and-rest) maker, simulated against the **real trade tape** (same
adverse selection a live maker eats), across 17 physics regions
(favorite/underdog × distance-to-strike in bps × realized-σ × seconds-to-close):

| region (favorite) | bid | P(win \| fill) | gap | edge_real |
|---|---|---|---|---|
| dist <20bp, ≤60s | 0.957 | 85.4% | −10pp | −0.089 |
| dist <10bp, ≤60s | 0.904 | 70.8% | −20pp | −0.179 |
| dist  <5bp, ≤60s | 0.704 | 44.2% | −26pp | −0.227 |

**Zero of 17 regions are positive.** A resting bid always fills ~10–20pp *below* its
true win probability — you are filled precisely when price is moving against you.
The rebate (hundredths of a cent) cannot close a 10–20pp gap. This holds near and
far from strike, early and late.

## 4. The cancel race — also closed

The passive maker is adversely selected everywhere, so the active maker's hope is to
**cancel an about-to-be-adverse quote before the informed taker hits it**. We simulated
exactly that on the microsecond book-delta firehose: post the favorite bid, fire a
cancel when the oracle-fair drops below the bid, and let the cancel land `L` ms later.
Sweeping `L` from 0 (perfect) to ∞ (never cancel):

| cancel latency | edge_real | $ROI |
|---|---|---|
| 0 ms (perfect) | −0.1898 | −42.4% |
| 190 ms (ours) | −0.1897 | −42.3% |
| ∞ (never cancel) | −0.1894 | −42.3% |

**The cancel does nothing — not even at 0 ms.** This is *not* a "we lose the latency
race" result; it is sharper. The cancel *signal* fires too late: the informed taker
hits the bid **before** the data-streams oracle reflects the move. The adverse flow
leads the oracle, so no oracle-based cancel can help at any latency. The only
untested variant is a flow-based cancel (detect the informed taker mid-arrival and
pull within the inter-trade gap) — which is the same sub-latency race a live attempt
already lost. **Maker is closed: passive (no region) and active (cancel useless).**

## 5. External replication (independent corpora)

- **Academic** ("Who Wins and Who Loses in Prediction Markets", 588M trades):
  ~70% of users lose; winners **provide liquidity with limit orders**, losers
  **take liquidity with market orders**; only ~12% of top winners beat a coin-flip
  benchmark and ~60% of "lucky winners" become losers out-of-sample.
- **Other people's bots** (a public arena of 4.77M paper trades, gamma-resolved):
  taker `edge_real −0.031` / ROI −6.0%; maker `−0.136` / ROI −8.9%. Even the maker
  families with positive `edge_real` still lose on realized ROI once fees and early
  exits are counted.

Both land where our own data does: the taker space is closed; the maker space only
opens on execution speed.

## 6. Takeaway for anyone using this dataset

Grade against the transactable price. If you find a +edge in a backtest here, before
believing it: (1) pool across coins to kill multiple-comparison luck, (2) require
the persistence lens to agree with the tape lens, (3) hold out by date, (4) check it
isn't a once-a-day pocket. We did, on everything, and the honest answer is that the
signal is already in the price.