Writing · Post-mortem
The Toll
Thirteen months, eight trading systems, 136 combinations measured under pre-registered criteria. Zero operable edge — and the verdict was written down before the last number arrived.
By Francisco · August 2026 · 5-minute read
Over thirteen months ending in August 2026 I built eight systematic trading systems for crypto and FX. All eight are dead. The difference is that the last of them died properly: measured against criteria committed to a repository before the results existed, by a framework whose job was to kill them if they deserved it. They deserved it.
That framework — internally, the Crucible — is not a backtester. It is the infrastructure that decides whether a backtest means anything: decision criteria committed to the repository before any result exists — the tooling refuses to run without them — thresholds derived from trading friction instead of chosen by taste, and 123 unit tests that validate every layer against series with known answers. Its function was never to confirm hypotheses. It was to make months-long mistakes die in days.
The arithmetic wall
The first finding isn't about any indicator. It's about arithmetic. For a signal to pay for itself, its information coefficient — the correlation between what it predicts and what then happens — has to clear a floor set by cost and volatility: ICmin = cost / σ(horizon). On real data, that floor is 0.383 at thirty minutes, 0.135 at four hours, 0.055 at one day, 0.021 at seven days.
An IC of 0.02–0.05 is what exists in nature. Above 0.10, in my experience, it is almost always look-ahead leaking into the test. Needing 0.383 at thirty minutes means no indicator was ever going to work there — not the six I built in that zone, not the next one either. There was no design error repeated six times. There was one framing error, inherited once: treating the intraday timeframe as a choice of style, when it was a choice of arithmetic.
Where the toll is payable, what appears is not edge
Push the horizon out and the floor drops to payable levels. Something does appear there — it just isn't edge. Every candidate that cleared the friction threshold turned out to be something else wearing its clothes.
Distance from the 252-day high looked like edge: +0.042 on the 40 symbols I had picked. On the 469 I hadn't: −0.013. It was measuring my selection, not the market. Twenty-day volatility had an IC that grew with the horizon — the signature of a risk premium, not a signal; the worst short in its sample rose +98,187%. Funding carry — collecting the periodic payments perpetual futures make to one side — was real, about +27.5% a year, with a single −87.9% month inside it. And the hedged version, cash-and-carry, removes the price risk while staying exposed to those payments: one ticker took −100% of the position.
The finding that appeared three times
The most useful result of the whole thirteen months fits in one sentence:
In crypto there is predictable structure, and it is systematically smaller than the toll.
Order-flow delta divergence: IC 0.016 — statistically genuine, under its friction floor. Apex's z-score: stability 1.00 across three horizons — under the floor. Open interest: IC −0.034, p < 0.0001, stability 1.00 — under the floor by seven times. Three times the signal was real. Three times the cost ate it. The market isn't random at these scales; it's tolled.
What a small sample actually does
The system that took longest to die took 98 days — and died without a verdict: nothing was decided, I just stopped looking. The last one, Apex, died in a single day, against 884 trades and a criterion written before I saw a number.
The contrast is the education. On its first eight exported trades, Apex showed 87.5% winners and +0.80R net — R being profit measured in units of risked capital. On 884 trades: 34.7% and −0.33R. Same indicator, opposite conclusion. A small sample doesn't give you a noisy version of the answer — it gives you the opposite answer, delivered with confidence.
What's left standing
The verdict closed a line of work, not the shop. The framework stands — it was never a bet on finding edge, so this result doesn't dent it. Along the way it caught data defects that would have minted false discoveries: a ticker discontinuity of 177,400× that pushed the friction threshold to zero, and a date-arithmetic bug quietly eating the last day of every range. It caught my own bugs with the same indifference. The data stands: ~1.4 million rows of validated crypto panels — spot, perpetuals, funding, positioning — with survivorship bias controlled. And the method stands: pre-register before results, derive thresholds instead of choosing them, and never confuse dying from lack of power with dying from lack of edge.
To be precise about what this verdict does not say: it doesn't say systematic trading is impossible, and it doesn't say crypto has no edge. It says these 136 combinations, on this data, at these costs, have none that survives the toll — and that the structure I did find was never bigger than the cost of capturing it.
Why publish a failure
Because the quiet alternative is the industry default. Eight systems died over these thirteen months, and the difference between this closure and the seven abandonments that preceded it is that this one has numbers, criteria written in advance, and a reason. The others were simply left behind — free to be remembered as "almost worked."
A documented kill is worth as much as a promoted success. This one is documented across 136 measurements, under criteria locked before the answer arrived. If we ever work together, that is the discipline your numbers get: the figure I hand you is the figure that's true — especially when it's the one I didn't want.