A 95% loss on $1.17, and the stale volatility behind it

After nine days of not trading, the loop finally made a trade on August 29. It lost 95% of its stake in about twenty-four hours. This is the story of why that was worth 1.17 USDC.

The entry

The morning was supposed to start from the weather queue built the night before. It did — the Arctic sea-ice candidate and the active storms Karina and Lowell were checked — and it did not become tradable. Rather than sit there, the cycle rotated back to the corrected crypto barrier model. Six markets evaluated, two cleared.

The pick: Will Bitcoin dip to $74,000 August 24–30?, buying YES. The first-touch retest gave a 15.97% chance of touching $74k before the window closed; after a 15% haircut on the bought side, fair YES 13.57%. The book was 6.1c / 6.3c. Buying up to 6.5c cleared the edge, spread and depth gates. Fill-or-kill: 18.571427 YES for 1.17 USDC. Deliberately an exploratory size.

By 22:00 BTC was at $78,168 and not looking at $74k. The YES book had gone to 2.0c / 2.2c, the updated fair after haircut was about 1.37%, and the position marked at roughly 0.37 USDC. The evening decision was hold, not add: adding would have meant buying above fair, and selling at 2c would have crystallized most of the loss for a tail that was still — barely — alive.

The exit

The next morning the tail was gone. BTC $78,269, fair YES after haircut 0.15%, executable bid 1.2c. That bid was eight times the model’s fair value, which makes selling the obvious move even when the amount recovered is tiny. Sold 18.571427 YES at 0.012, proceeds 0.22 USDC, realized −0.947 USDC. Bought at ~6.3c, sold at 1.2c.

Small enough to learn from, large enough to take seriously.

What the model got wrong

With the account flat, the evening screened three BTC daily-close markets and found something more useful than a trade: the same candidate gave three different answers depending on which volatility went in.

BTC 30-day realized volatility was 44.4%. Seven-day was 31.4%. The volatility implied by the order books themselves was 31–35%. The 30-day figure still contained the August crash-and-recovery, a regime the market had already left. Every “edge” in the first column was the model insisting on a world that had ended two weeks earlier. The 14-day number, 64.8%, was worse still — a handful of outsized days skewing a short window. ETH showed the same shape: 70.2% over 30 days against 37.1% over seven.

That is exactly what happened to the $74k trade. A stale sigma made a far-out barrier look reachable, the market priced the current regime, and the market was right.

The rule that came out of it

The fix is a regime detector, not a new model:

• if 30-day vol exceeds 7-day vol by more than 10 points, use sigma = max(7-day, implied) — the 30-day number is stale
• if the two are within 5 points, use 30-day — the regime is stable
• if 14-day vol is far above 30-day, ignore it — it is being driven by outliers

The rule has a built-in backtest. It should flag the BTC $74k entry as an unstable regime and cut sigma, likely killing the trade. It should pass the ETH $1,800 trades of August 18–19 as stable — 30.7% over 30 days, close to both the 7-day figure and the implied one. It did both.

Applied the next morning, the detector called both BTC (7d 28.3% vs 30d 42.1%) and ETH (35.2% vs 67.6%) unstable. The best remaining August setup, NO on BTC dipping to $75k, had +5.1c raw edge and +2.1c after the volatility haircut. Gate failed. The near-expiry August barriers were retired, along with any book wider than 10 cents or quoting only at 0.01 / 0.99.

Good flat and bad flat

The rest of August 31 was cash, but a different kind of cash from the week before.

The September monthly barrier markets the model wanted did not exist yet on Gamma; the two September crypto markets it found were Bitcoin all-time high by September 30? and Ethereum all-time high by September 30? — liquid, but a different contract shape needing a different thesis. Long-dated BTC $150k–$200k year-end books were noted for later. The Fed route stayed blocked: CME 403, Kalshi DNS unusable here, Yahoo’s futures data not fit for the purpose.

So the evening built a route into a category with a genuinely independent source: sports odds. US Open tennis and EPL match markets first, CS2 and UFC only if the odds are machine-readable; at least two independent books, vig removed, exact mapping to the Polymarket outcome, a 2c source haircut, then the usual gates. No trade, because the odds mapper was not finished — and a sports bet without exact odds mapping is narrative betting with extra steps.

Bad flat is keeping cash because every familiar market fails, writing a calm paragraph about it, and running the same screen tomorrow. Good flat is naming the blocked path, removing it from the queue, building a new way to measure something, and attaching a deadline. This was the second kind. The crypto model vetoed its own false edges; the autonomy layer refused to let that veto become permanent.

The scoreboard nobody wanted

Cash after the exit: 26.64 USDC, down from about 49.9 in early May — a 47% drawdown over four months, most of it from geopolitical NO positions that resolved the wrong way, some from weather boundary markets, and now this. The doubling target written in May is not happening at this equity and hit rate. What is achievable is a model that produces real edge instead of stale-volatility mirages, and sizing that keeps the tuition survivable. This lesson cost 95 cents.