This is engineering research about the absence of edge, published as method. Nothing here is financial advice, a signal, or a recommendation to trade.
Abstract
Over nine weeks we built AxT, a validation-first trading research harness, and used it to test twelve classic systematic strategies — trend-following, mean-reversion, session breakout, carry, statistical arbitrage — across FX, gold and crypto, on up to eleven years of data. None passed our pre-committed statistical gate. We are publishing the full result set, including one hypothesis invalidated by our own engineering bug, one real edge that decayed below viability, and the deliberation process between the human operator and the AI research partner that shaped the project's direction. A thirteenth hypothesis is pre-registered and untested; the project is deliberately paused until we have better evidence about where to look next.
1. Why build a machine designed to say no
Retail trading content overwhelmingly publishes survivors. Backtests that worked, strategies that fit one lucky period, screenshots without sample sizes. Our first-generation system was part of that problem: it ran a sophisticated multi-model deliberation pipeline over a signal that had never been backtested at all. In 24 hours of observation, the deliberation layer rejected 27 of 27 signals — at real API cost, with zero measurable contribution. The post-mortem of that failure produced the founding rule of the second generation: no strategy earns forward-testing, let alone capital, before proving statistical edge on years of data — and the validation machine must be built before the strategy.
2. Method
The harness (5,872 lines of Python, 151 commits, 104 automated tests) enforces discipline structurally rather than by good intentions:
One contract, two worlds. The exact same strategy code — propose(candles, state) → Signal | None — runs in backtest and live simulation. A whole class of silent divergence bugs disappears by construction.
No look-ahead, by construction. The simulator only shows a strategy candles up to the decision bar, and fills every decision at the next bar's open. Costs (spread, slippage) are applied by the engine, never inside strategy logic, so no strategy can quietly forget them.
Honest position sizing. Sizing uses true fractional risk with a real leverage ceiling. The code documents why the simpler alternative — a fixed lot cap — was rejected: it would have silently converted fractional risk into fixed lots and understated drawdown, letting strategies pass the risk gate spuriously.
A gate with no "almost". A strategy passes only if annualized Sharpe > 1.0 and profit factor > 1.3 and max drawdown < 15% and the sample has at least 100 trades. There is no discretionary override, and results must hold year-by-year — an eleven-year average is not allowed to hide a dying edge.
3. Results: twelve tested, zero passed
Eight clean failures in FX and gold: Donchian trend-following in two variants (Sharpe as low as −1.35, drawdowns to −79%), Connors-style RSI(2) mean-reversion (benign drawdowns, but profit factors below the gate), a stop-distance sweep proving the stop was never the problem, canonical 5/95 mean-reversion (insufficient sample — reported as such, not massaged), London session breakout (clean fail), and the canonical Turtle system on daily gold — our best trend result at PF 1.03 and +1.6% over eleven years, which is noise, not edge.
One inconclusive: Donchian on BTC, invalidated by our own integer-sizing bug that rejected fractional positions. We report it as inconclusive — treating it as a clean FAIL would overstate what we learned, in the flattering direction.
Three crypto failures, each instructive:
- •Cash-and-carry (spot + perpetual funding). The most interesting result of the project: a real, measurable edge that existed — roughly 30.6% annualized in 2021 — and decayed to about 5.1% by 2025 as the market competed it away. This is categorically different from "no edge": it is direct evidence of how quickly public edges die. We publish it strictly as a decay study.
- •BTC–ETH statistical arbitrage. The pair is no longer cointegrated (p = 0.24). The thesis wasn't misimplemented; it became structurally false after the ecosystem diverged in 2022.
- •Grid + funding compounding. Superficially profitable (+$22.9k over five years) — until you decompose it: +$205.7k of winning rotations and +$11.9k funding against −$194.7k of losing rebalances, over a period in which the underlying rose ~200%. A short-gamma signature dressed as income.
One number about numbers: an earlier internal draft of this project claimed "15 hypotheses tested." The project's own decision document corrected it — "rhetorical inflation, corrected" — to twelve. If a research lab's discipline doesn't survive its own marketing instincts, it isn't discipline.
4. The partnership pattern
The project's direction was set by documented deliberation between the human operator and the AI research partner. After nine FX/gold failures, the AI recommended pausing the project. The operator pushed back with a legitimate, testable objection: classic edges may survive longer in less efficient markets — test crypto before concluding. The AI assessed the objection as good science rather than attachment, executed the pivot, and when all three crypto hypotheses also failed, updated its pause recommendation from "probably right" to "strongly supported" — and the operator then pre-registered one final hypothesis with hard stopping criteria the AI helped specify. Disagreement, evidence, updated positions, pre-commitment. That is what "AI as a partner" means in this lab — not a tool that obeys, not an oracle that decides.
5. Limitations
No strategy passed, and no real capital was ever deployed — the entire record is backtest and demo. The hypothesis space we tested is the classic, documented, individually accessible one; we make no claim about machine-learning regimes, order-flow microstructure, cross-exchange arbitrage or options structures, which remain untested here because their infrastructure cost is months, not days. One test was invalidated by our own bug. And a single operator's harness, however disciplined, is still a single implementation — independent replication would strengthen every number above.
6. Status and what's next
The thirteenth hypothesis — flow and positioning signals (funding as signal, open-interest deltas, liquidations) — is formally pre-registered with a five-week budget and hard stopping criteria. It has not run. The project is deliberately paused: not because the question is answered, but because choosing the next test well requires more evidence about other markets than we currently have. The harness, the gate and the method are preserved as reusable engineering for whatever systematic question comes next.
Twelve hypotheses entered. Zero survived. The machine that killed them is the most valuable thing we built.
Research home: blackicelabs.ca/research · This note is engineering research, not financial advice.