This is engineering research about market efficiency and validation discipline. Nothing here is financial advice, a signal, or a claim that any of it makes money — the finding is the opposite.
Abstract
We built a research system to answer one question: is there a capturable edge in binary prediction markets for a small retail operator — either by predicting outcomes better than the market, or by capturing structure (market-making, arbitrage, liquidity rewards)? We pre-registered a validation gate before looking at any data, ran the system in shadow mode, and got a headline result: an 80% win rate. The same gate then proved that number was noise. This note is about that moment, because it is the whole reason the discipline exists.
1. The setup
The system was the sibling of our trading harness (see Lab Note 002), pointed at a categorically different market: binary, single-resolution contracts instead of continuous price. That difference mattered enough that we couldn't reuse the old gate — Sharpe ratio and drawdown don't describe a market that resolves to exactly $0 or $1. So we wrote a new one: eight conditions, all required, covering sample size, edge magnitude, probability calibration, and — critically — one anti-survivorship condition borrowed from a lesson the trading harness taught us: no single month may account for more than 60% of total profit. That condition was written before any data was seen, with an explicit note in the document: any reclassification after seeing data is search-noise.
2. The 80% win rate, and why it was a lie
Then the audit hit a wall of our own making: a destructive database command wiped the production results, no backup. What survived was the log — roughly 25,000 shadow decisions with enough detail to reconstruct outcomes, because these particular markets resolve deterministically from public spot price. So we rebuilt the result set from the log against free public price data, and measured: 80.3% win rate, +25% gross ROI.
That is a number you put in a headline. The gate required a confounding check anyway, so we ran it — and it dismantled the result. Bitcoin fell 19.3% during the measured window. A "NO" bet on a strike market is structurally a short position on Bitcoin; with Bitcoin falling, those bets won almost everything — 94% on the bearish side. The "YES" side, the half of the strategy not helped by the drop, won 55% and netted essentially zero. The edge wasn't prediction. It was directional beta, coincident with one month of a falling market — exactly the regime-concentration trap the anti-survivorship condition was written to catch. The strategy had no predictive skill. The validation, in the report's own words, paid for itself.
3. The second question, killed cheaply
Predictive edge dead, we asked the other half: is there structural inefficiency to capture — spreads, combinatorial arbitrage, liquidity rewards? Here the honest move was methodological: we pre-registered this as a disconfirmation test, not a confirmation test. Passive data can prove market-making won't work (spreads never wide enough, no real depth); it cannot prove it would, because filling a market-maker's order is itself adverse selection. A scan of 1,067 live-orderbook markets settled it fast: liquid markets (>$100k) show a median spread of 0.1¢ — bot territory. The illiquid tail has wide spreads but $1–50 of depth and ~$200/day of volume, and where real depth exists, the counterparty is a sharp sports bettor. Liquidity rewards pay you precisely to hold adversely-selected inventory, and dilute as more providers arrive. No capturable structure at retail scale. Stop.
4. What it cost, honestly
The real bankroll was $21.19 — the $1,000 configuration was never funded, and the pivot document treats that as a hard limit, not a footnote: a small edge is an upper bound on a large one, because market impact and adverse selection only get worse with size. Along the way one legacy module leaked $11.57 through a shared live_trading flag read by the wrong engine — a real, small, quantified defect, disabled and documented rather than buried. And the destructive-command data loss is its own lesson: the method survived (log reconstruction worked), but the absence of a backup routine is flagged in our own report as the first thing to fix before any re-run.
5. Why publish a system that made nothing
Because the deliverable was never the system. It was three things: a validation discipline that demonstrably caught its own false positive before any capital or conviction rode on it; a working system with an operational spec, preserved; and an empirical map of what does not work in retail prediction markets — predictive edge (regime beta in disguise) and structural capture (adverse selection all the way down). The verdict was pre-committed and self-triggered: STOP. That is what a rigorously tested negative looks like, and a lab that only publishes its wins isn't running experiments — it's running ads.
6. Status
Archived. The research question is answered in both forms it could take, by two independent methods that agreed. The system is dormant; the method — pre-register the gate, distrust the flattering number, kill cheaply before building expensively — carried directly over from our trading research and will carry into whatever comes next.
The win rate was 80%. The edge was zero. Knowing the difference is the entire job.
Research home: blackicelabs.ca/research · Engineering research, not financial advice.