Koryu · 黄龍
Koryu
← Research library
Research · 研究 · The system

We found a way to make the short work. Our own validation gate told us to slow down.

21 Aug 20266 min readThe systemKoryu Research

Nothing here tells you what to buy or sell. This piece reports how we improved one part of our candidate strategy, why our own validation refused to fully bless it, and what we did about that. Nothing here is investment advice. Decisions are yours.

We have written that on this market the edge is the regime, not the coin. The system built on that idea has a quiet second half: a short side that only wakes up when the market turns down. It was the weakest part we had. This is the story of rebuilding it, including the moment our own test told us to stop admiring the result.

The half that acts in the worst weather

Our candidate strategy is long when the market is risk-on, short in a confirmed downtrend, and in cash the rest of the time. The short exists to do one job: earn something when everything is falling. For a long stretch it did that job badly. On its own it dug the whole system's worst-case loss almost ten points deeper than simply stepping aside into cash would have, and our honest conclusion, more than once, was that the short did not earn its place. We came close to cutting it entirely.

Two unglamorous fixes: patience, and taking the money

Before changing anything, we fixed how we measured. A long sleeve and a short sleeve cannot be tested in separate spreadsheets and then stapled together, because in real life you hold one account and the two trade off against each other. So we rebuilt the test as a single account that switches between long, short, and cash, closing whatever it holds the moment the regime flips. Judged that way, on data that includes the coins that went to zero, the old short looked even worse than we thought. Good. An honest ruler is the whole point.

With that frame, two plain ideas moved the needle. The first was patience. Instead of shorting the instant the market slipped below its trend, we made it wait for the downtrend to hold for several days first. Crypto crashes are violent in both directions, and the opening days of a drop are exactly where the vicious relief-bounces live; waiting stepped around most of them. The second was taking the money: once a short had fallen a set amount, close it and bank the gain rather than ride it into the snap-back that so often followed.

Neither idea is clever. Together they pulled the short's extra drawdown back to roughly the level of just holding cash, and lifted its risk-adjusted return above the do-nothing baseline. For the first time in the whole investigation, the short looked like it belonged.

Then we tried to break it

A good backtest is a hypothesis, not a conclusion. So we ran the result through the same gate every candidate here has to pass. Train the choice on old years and grade it on years it never saw; then, separately, estimate the probability that the winning settings were simply the luckiest of the many we tried.

Two of the three checks came back clean. Trained on the past and graded forward, the improved short beat the cash-only baseline, and in the 2022 bear it turned what had been a losing year into a winning one. Deflated for the number of variations we had tried, its edge still cleared the bar. The effect is real.

The third check is the one that matters. When we asked how likely it was that the precise recipe we had landed on, wait this many days, take profit at that level, was genuinely the best rather than the luckiest, the answer came back a little worse than a coin flip. Every near neighbour was about as good. We had found a real effect and a fragile address for it.

A fragile setting is a stop sign, not a green light

This is the least intuitive and most important part. The improvement is real: patience and profit-taking help, and they help across a whole family of nearby settings. What is not real is the exact number. If we shipped the single best-looking configuration, we would be doing the one thing we built Koryu to refuse, mistaking the luckiest point on a noisy surface for a discovery. The honest response to "the direction is real but the precise dial is noise" is to step back from the peak to a calmer, more conservative version, and to expect less of it.

So we shadow it. We do not adopt it.

We picked a deliberately patient version of the improved short, not the flashiest one, and put it where every unproven idea here goes: a forward, public shadow record that logs what it would have done, day by day, with no money behind it, live on the shadow page. If it earns its keep in real time, on markets it has never seen, we will consider promoting it. If it does not, cash was the better answer all along. Nothing we actually treat as the strategy of record changed, because none of this is traded.

The satisfying version of this story ends with a big number and a victory lap. The honest version ends with a real improvement, a validation gate that made us slow down, and a conservative bet placed in the open for the market to judge. That is slower and far less glamorous, and as far as we can measure, it is how you keep from fooling yourself. What you do with it is your decision.

Related reading
The systemThe real edge is the regime, not the coin6 min readThe systemThe capitulation floor: implied volatility joins the gate7 min readThe systemStrategy v2.0 in shadow: the full backtest, before we trust it9 min read
Frequently asked

What was wrong with the short side of the strategy?

On its own it deepened the whole system's worst-case drawdown by nearly ten points compared with simply sitting in cash during downturns. For a long stretch our honest read was that it did not earn its place, and we came close to removing it entirely.

What changed to make it work?

Two plain adjustments, tested in one combined account rather than as a separate sleeve: waiting for a downtrend to hold for several days before shorting, and taking profit once a short had fallen a set amount instead of riding it into the bounce. Together they pulled the extra drawdown back toward the cash-only level and improved risk-adjusted return.

If the backtest improved, why is it not live?

Because our validation gate only partly blessed it. Trained on old data and graded on unseen years it beat the do-nothing baseline, but our overfitting check put the odds that the exact settings were the genuine best, rather than the luckiest we tried, a little worse than a coin flip. A real effect at a fragile address is a reason to slow down, not to ship.

What does it mean to run the short in shadow?

A shadow is a forward, public record of what a strategy would have done, computed every day with no capital behind it. The improved short runs there in a deliberately conservative version so it can prove itself on markets it has never seen before anything is adopted. The strategy of record is unchanged.

Does this change what Koryu publishes today?

No. The measurements and the regime read are unchanged, and nothing here is traded. This is a note about how a candidate improves under a validation gate, published in the same spirit as the rest of the site: the process in the open, including where it told us to wait.