← Research library
Research · 研究 · 55 · Engine

Inside Seiryū. The recovery engine, by the numbers.

12 Aug 20267 min readMethodologyShishin Research

This article describes one of Shishin’s research engines and the results it produced in a hypothetical five-year backtest. Backtest figures are not live trading results and do not predict future performance. It is educational, an explanation of how a systematic engine works, not personalised investment advice, and not a recommendation to buy or sell any security.

Seiryū (青龍) is the most surgical of Shishin’s four engines. It was the active engine on just 42 of the 1,258 backtested days, fewer than any other, and it took only 24 trades in five years. From that sliver of the calendar it produced 10.3% of the entire stack’s gains, the highest return per day on duty of the four. It is the engine the system keeps holstered almost all the time, and draws only at a very specific moment.

When Seiryū is allowed to trade

Markets do not exit bear regimes cleanly. They transition messily, with false starts, head fakes, and a few weeks where large, liquid names lead a bounce that has not yet broadened into the small-cap universe. That narrow window is the only thing Seiryū is built for. The regime classifier deploys it when the read is, roughly, the bottom may be in, but it is the big names leading: too early for the breakout engine to earn its keep, but the wrong moment to sit entirely in cash. Most regimes never call for it, which is exactly why it is on duty so rarely.

What Seiryū looks for

Inside that gate, Seiryū hunts speculative recoveries among more liquid names than the small-cap engines touch, the kind of stock that can lead a turn rather than follow it. It ranks candidates with the system’s composite score, leaning on the setup-state classifier to prefer names that are turning up out of a base rather than ones that have already run. The backtested book reflects it: its largest winners were recovery names like OUST (+76%), TSSI (+66%), and SMCI (+47%), bounces caught in the transition window rather than chased at the top.

How it enters and exits

Seiryū fills on the close and manages risk with the same two-sided exit the momentum engine uses: a position is held until it closes below its 12-day moving average or hits a volatility-scaled stop, 1.25× the 14-day ATR, floored at 2%. In the backtest the split was almost even, 13 of the 24 exits were the trailing moving average and 11 were the stop, and the median holding period was about 12 trading days. The recovery thesis gets a normal leash, not a special one: if the bounce does not hold, the trade is cut on the same terms as anything else.

What its trades actually look like

Here are all 24 of Seiryū’s backtested trades, binned by outcome. With so few trades the distribution is sparse, but its character is clear:

048−20−10−50+5+10+20+40+80per-trade return (%)median −0.8%mean +9.1%12 losers12 winners · long tail →
Fig., Seiryū’s 24 recovery trades, by outcome (locked five-year backtest, hypothetical). An even 12-and-12 split of winners and losers, with the winners running further than the losers fall. Sparse by design: this engine trades rarely. Past performance does not predict future results.
  • 50.0% win rate. 12 winners, 12 losers, the most balanced hit rate of the four engines.
  • The median trade was roughly flat (−0.8%), a near-scratch typical outcome rather than the clear small loss the momentum and quality engines show.
  • The mean trade made +9.1%: the average winner (+26.0%) was a little over three times the average loser (−7.8%). The winners are tighter than the small-cap engines’ (whose biggest names ran past +200%), which is what you would expect from more liquid recovery leaders, the worst trade, QUBT, lost only 11.9%.

Deployed rarely, on purpose

The striking number is not any single trade, it is the ratio. Seiryū was on duty 42 days and contributed about $770,000 of backtested profit, the best gain-per-day-on-duty of the four engines. That is the whole design: an engine that does nothing for long stretches, then is deployed precisely into the short window where a recovery is real but small-cap momentum has not yet earned the right to fire. An engine like this would be impossible to run on its own, you would spend years waiting, but as one of four it can afford to wait, because the other three are covering every regime it is not built for.

What we tested and didn’t keep

Seiryū’s rarity invites two obvious upgrades, and both were rejected. The first was a dedicated recovery-scoring model to grade its candidates more aggressively. It looked good in isolation and got worse once placed inside the live regime-gated system, a clean example of a standalone improvement that does not survive contact with the rest of the stack, so we did not adopt it. The second was flattening its position sizing rather than letting the composite size each name; it hurt returns and was dropped. Both are catalogued with the rest in the experiments that failed, and both reinforce the house view: the gains come from deploying the right engine at the right time, not from over-engineering any one of them.

Why Seiryū is one of four, not the whole strategy

A 24-trade engine that only works in the transition out of a bear market is, by itself, not a strategy at all. Its value is entirely contextual: it fills the seam between the defensive engine that works while the broad market is falling and the breakout engine that works once the recovery broadens. Hand that seam to a momentum engine and it trades too early; hand it to cash and it misses the leadership entirely. Seiryū exists to own that one transition, and nothing else, which is the logic of four engines for four regimes.

Sources & methodology

The figures here are the Seiryū slice of Shishin’s locked five-year backtest (hypothetical), counted per position with the late-stage partial trims merged back into their parent trade. How the book is measured, survivorship-free universe, close-fill convention, the leave-out-the-winners robustness test, and the per-trade significance battery, is documented in the five-year record, leave out the winners, why backtests lie, and statistical significance. The other engines are profiled in inside Suzaku, inside Genbu, and inside Byakko, and the framework that switches between them is four engines for four regimes.

Related reading
MethodologyMarket-on-close: why we place every entry at the close, not the open7 min readMethodologyWhy most backtests lie: overfitting, look-ahead, and the walk-forward fix10 min readMethodologyThe setup state classifier: context beats raw signal8 min read
Frequently asked

What is Seiryū?

Seiryū is Shishin's recovery engine, one of four regime-specialist engines. It hunts speculative recoveries among more liquid names during the messy transition out of a bear regime, when the bottom may be in but large names are leading and small-cap momentum has not yet earned the right to fire.

Why does Seiryū trade so rarely?

It only fits one narrow regime, the transition out of a bear market, so the macro classifier deploys it on very few days: just 42 of 1,258 in the backtest, the fewest of the four engines. From those it produced about 10.3% of the stack's gains, the highest return per day on duty.

What was Seiryū's win rate?

In Shishin's hypothetical five-year backtest Seiryū won 50% of its 24 trades, the most balanced hit rate of the four engines, for a mean of +9.1% per trade. Its winners were tighter than the small-cap engines' because it trades more liquid recovery leaders. Past performance does not predict future results.