The Proving Ground

Locked 2026-11-05 01:49 UTC

Question. When most accounts on a coin are short, does the coin outperform over the next 72 hours?

Bar to pass. t ≥ 2 on 2023-09 onward, earlier data the same sign, after fees and funding.

Result. Passed its bar: t 2.31 on 2023-09 onward, earlier data the same sign (+1.10%). The average trade made +1.01% after fees and funding, but the typical trade lost 1.18% and only 45% won. The average comes from rare squeezes of up to +910%. Without the best 5 weeks, t falls to 1.08.

What we learned. Our first pass in ten tests. One pass in ten is close to what luck alone produces, so this earns more testing, not money.

What we did. The forward trial continues to its locked verdict. Nothing is scaled until that confirms it.

Lock file · sha256 e3a5e554114b9a62…

Locked 2026-11-05 00:21 UTC

Question. After Bybit announces it will delist a coin, does that coin keep underperforming the market for a week where it still trades?

Bar to pass. t ≥ 2 on events from 2024-09 onward.

Result. Failed (t −0.85). The typical coin did keep falling (median +3.3% on the hedged short, 65% of the time), but a single delisting squeeze sent one coin up about 30× and erased everything.

What we learned. Shorting weak coins wins often and loses enormously. That is selling insurance against squeezes, not an edge.

Lock file · sha256 d7717fa66d691d67…

Locked 2026-11-04 23:30 UTC

Question. If a Telegram signal channel was significantly wrong in one period, does betting against it pay in the next?

Bar to pass. Rank channels on May–August 2026, trade against the worst on September–November, t ≥ 2 after fees.

Result. Failed (t −0.56). Only one channel was significantly wrong in the first window, and it was not wrong in the second. A channel’s rank in one window told you nothing about the next (correlation +0.12).

What we learned. Once you subtract what the market did, the channels we track have no stock-picking skill in either direction. Their results are mostly market direction. That is now a column on our panel.

Lock file · sha256 f7eff75c63a8518a…

Locked 2026-11-04 23:23 UTC

Question. If the rebound after a BTC crash comes from forced selling, baskets whose open interest was flushed harder should rebound more. Do they?

Bar to pass. Rank correlation below zero with p < 0.05.

Result. Weak (correlation −0.09, p 0.11). Right direction, not established, and nothing in the recent half. Most crash days had no open-interest flush at all.

What we learned. The crash rebound may be real, but this explanation for it is not proven.

Lock file · sha256 1568b8678ad0629a…

Locked 2026-11-04 23:17 UTC

Question. After BTC drops sharply, do the coins that fell most bounce back over the next 24 hours?

Bar to pass. t ≥ 2 on 2023-09 onward, and still t ≥ 2 with an extra 0.3% trading cost on crash days.

Result. Inconclusive: +1.21% per basket, t 1.86 in the recent half (t 3.40 before it). With the extra cost, t 1.40. Our own trial fills on a crash day cost about 0.02%, so cost is not the problem; the small number of crash events is.

What we learned. Our strongest lead, and the only idea here whose wins are larger than its losses. Not proven yet.

What we did. Our forward trial keeps running, unchanged, to its pre-set 15-basket verdict.

Lock file · sha256 4ddb014b2e1dff05…

Locked 2026-11-04 23:01 UTC

Question. When a coin’s futures price drops below its spot index while BTC is falling, does it rebound over 24 hours?

Bar to pass. t ≥ 2 on 2023-09 onward, earlier data the same sign.

Result. Inconclusive: +0.32% per trade, t 1.65. Almost all of it is the general bounce after BTC drops; the discount itself adds little.

What we learned. We first scored this as a pass (t 3.79) by averaging per day. Scored per trade, as the lock file specified, it is not. We publish the correction.

What we did. Our forward trial keeps running to its pre-set 300-trade verdict.

Lock file · sha256 a7f0f5d6c3f9d947…

Locked 2026-11-04 23:00 UTC

Question. The same short as below, but exiting if the coin rises 50%. Proposed after seeing the result below, so exploratory only.

Bar to pass. t ≥ 2 on 2023-09 onward.

Result. Shelved (t 0.79). Every year from 2020 to 2024 was positive; 2025 and 2026, with a flood of listings, were negative.

What we learned. It looks like a known trade that got crowded.

Lock file · sha256 0985183c368edcad…

Locked 2026-11-04 22:59 UTC

Question. Do newly listed perpetual futures underperform the market over their first month?

Bar to pass. t ≥ 2 on listings from 2023-09 onward.

Result. Failed (t 0.28). The typical new coin did fall (median +20% on the hedged short, 71% of the time), but squeezes of up to 37× wiped it out.

What we learned. A high win rate and a positive median can still lose money. Judge the average.

Lock file · sha256 0f32e83820baa22a…

Locked 2026-11-04

Question. Our main strategy bought the strongest coins and shorted the weakest. Its backtest scored t 2.36, but it only included coins still listed today. Does it survive once dead coins are included?

Bar to pass. Keep it if t ≥ 2; stop it if t < 1.5. Rule set before the run.

Result. Failed: t 0.09 to 1.02 depending on the liquidity screen, and under 2 even with zero fees. Most of the old result came from leaving dead coins out.

What we learned. Survivorship bias made our best backtest look real, and it had been losing in forward trading.

What we did. We stopped the strategy and closed every position on 2026-11-05.

The keep/stop rule was written in our research notes before the run; there is no separate lock file for this one.

Locked 2026-09-26 06:04 UTC

Question. The one channel that looked skilled on past data: does its edge hold on calls made after we locked the test?

Bar to pass. t > 2 on at least 50 independent calls over at least 30 days, measured 72 hours after each call.

Result. Rejected: 942 independent calls over 37 days, −1.94% per call, t −6.57.

What we learned. The best channel on past data was among the worst going forward. Past performance selected for luck.

sha256 4be52aad55dd805a… · Lock file held privately (read-only since 2026-09-26) because it names the channel. Its hash is shown so it cannot be changed after the fact.