Shadow Edge — R&D Status

1. Purpose of this document

Shadow Edge is a research program, currently running as a live real-money pilot at deliberately trivial size. This page describes it as factually as possible — including the failures — for readers with a professional quantitative background whose criticism is welcome. It deliberately avoids marketing language: where a number is a hypothesis rather than a fact, it is labeled as such. We publish the bad days alongside the good ones because a demonstration that omits them demonstrates nothing.

2. The strategy, literally

Polymarket runs daily binary markets on BTC and ETH closing prices ("will BTC be between $X and $Y today", "will BTC be above $X today"). The prices of those contracts frequently diverge — often by double-digit percentage points — from the probabilities implied by the Deribit options chain for the same expiry. Shadow Edge sides with the options market against the prediction market whenever the divergence exceeds a threshold, and holds to same-day resolution. There is no overnight exposure by construction, and no claim of arbitrage: it is a relative-calibration bet that a deep, professionally market-made options venue prices short-dated distributions better than a thin, retail-driven prediction venue, by more than the cost of trading the latter.

A complementary duration-matched hedge — opposite-side Deribit option spreads against Polymarket "above" positions, with the option legs closed at a fixed time so effective duration matches the same-day expiry — has been designed but is unfunded and untested.

Intent. The crypto-binary book is the first application, not the destination. The project is being open-sourced partly as a live, verifiable demonstration of capability. The ultimate goal is to use smart money — the professional odds embedded in derivatives markets — to price stocks, bonds, commodities (including cryptocurrencies), and prediction-market contracts accurately across a variety of venues. Polymarket crypto binaries are simply where the divergence is largest and the feedback loop is fastest. This is just the beginning.

3. The record to date, unvarnished

The strategy is being tested with real money at deliberately trivial size, and the test so far has been dominated by operational failure rather than by the signal.

Calibration evidence is thin: on an early 12-trade paper sample the model was conservative in the 50–70% implied range and overconfident in the ≥80% tail (95% implied realizing about 75%). The July entry-window amendment partly exists to keep the book out of that tail until it is understood.

4. What looks promising

First, the divergences are real, persistent, and large — multiple contracts on a typical day priced 5–25pp away from the options-implied probability, recurring daily rather than as one-off dislocations.

Second, the one cohort traded inside the intended discipline — the 55–75¢ entries — has a perfect record over an admittedly small sample, and that band is now the entire entry window. Every documented loss to date traces to a trade the written rules would have rejected. That is either an encouraging fact or an elaborate form of hindsight overfitting; both readings are held open.

Third, the infrastructure now exists end to end: automated signal generation from live Deribit chains, server-side limit-order execution, automated exits, settlement, and — as of this month — a required rejection ledger so declined trades can be audited.

Fourth, the capacity constraint cuts both ways. The book is far too small for institutions to compete over, which is precisely why the inefficiency may persist, and why the scanner has been open-sourced — partly as defence, partly as the live demonstration of capability described in §2.

5. What remains unsolved

Statistical proof. No sample yet distinguishes edge from luck. Sixteen disciplined wins is suggestive and nothing more. The current review trigger is 30 resolved trades under the new entry window.

Production parity. The recurring failure mode has not been the model but the gap between specification code and what the live runner enforces. The July amendment requires production to execute the strategy file verbatim and publish its rejection ledger. The technical team is now three — a principal, a developer, and a dedicated verification lead whose role is to check one side's work against the other's — providing an independent check absent when the June losses occurred. Continuous spec-to-production verification nonetheless remains an open control problem.

The measure problem. A Deribit-implied density is risk-neutral; no claim is made of recovering true real-world probabilities. The working position is that pricing off professional options market makers beats pricing off retail prediction-market flow; the 0.85 linear haircut is an ad-hoc correction chosen directionally from twelve data points.

Correlated tails. Same-day, same-direction BTC and ETH positions are close to one bet. They are capped under a shared budget and overlapping losing regions are rejected, but the joint tail has not been properly sized.

Data completeness. Deribit chain capture was disabled for a stretch, leaving a gap in the regression record. As of this week, every live signal is to be archived with the timestamped option chains from which it was derived.

Credential security. One area where the partitioning is reasonably clean: the runner authenticates to the venue with API credentials that can place and cancel orders and read account state, but have no withdrawal capability — fund movement requires a separate OTP-gated login, and custody sits in a proxy contract wallet whose owner key is kept offline behind 2FA rather than on the execution box. The residual risk of leaked trading credentials is order abuse (including self-matched mispriced limit orders), which is bounded by the account balance and accepted at pilot size.

Capacity. Even on favorable resolution of all of the above, per-market liquidity caps the current book at boutique scale. This is framed deliberately: the crypto-binary strategy is the demonstration, not the product. The capacity of the broader program — applying derivatives-implied pricing to stocks, bonds, commodities, and prediction contracts across venues — is a separate and much larger question. What the demonstration is worth if it cannot itself scale remains a fair challenge, and is posed in §6.

6. Open questions we are asking

These are the six questions we are actively putting to quantitative practitioners, in descending order of urgency. Considered answers to any of them are welcome.

1. Test design. Given tiny stakes and one-day cycles, what is the cheapest credible test that realized frequencies beat entry prices by more than costs — a fixed trade count, coverage across volatility regimes, or something else? At what sample size does "luck" stop being the better explanation?

2. Measure change. Is there a defensible middle ground between using the risk-neutral density raw and a naive linear haircut — for example, a variance-risk-premium-style adjustment appropriate to daily horizons — or is any transformation decoration at this sample size?

3. Correlated-tail sizing. How should a book of same-day crypto binaries be sized when the losing regions correlate to one underlying move? Even a rough framework would be valuable.

4. Controls. What is the minimum viable version of institutional change control — spec-to-production verification, pre-deployment checks, kill criteria — that a three-person technical team (a principal, a developer, and a dedicated verification lead) can actually sustain? The losses to date are almost entirely a controls story, and the verification role was added precisely because of it.

5. Calibration. With small n, how should the implied-versus-realized analysis be structured so it informs the haircut without overfitting?

6. Capacity, reframed. The current book is treated as a live, open-sourced demonstration of a broader thesis: that professional derivatives pricing can be used to price stocks, bonds, commodities, and prediction contracts accurately across many venues. Does a small but fully verifiable live track record persuade sophisticated allocators of a methodology — or is it discounted to zero unless the demonstrated strategy itself has capacity? If the former, what would make the demonstration most convincing?

Contact: Nicholas Levenstein — nick@levenstein.net.