The figures in this article come from a backtest of Shishin over 1,258 US trading days (2021-05-12 → 2026-05-14). Past performance does not predict future results. Shishin is a research publisher; published material is research output, not personalised investment advice.
A quantitative strategy that works in one regime is a research artifact. A quantitative strategy that survives every regime is a capital allocator. The difference is whether you've built one model or four.
Shishin is a four-engine US-equity momentum system: four independent strategies, Genbu, Suzaku, Byakko, and Seiryū, each built for a different market regime, with a daily breadth classifier deciding which one is allowed to trade.
The premise
Almost every long-bias equity strategy you can buy off the shelf is built to work in a single market state. Momentum thrives in trending markets and bleeds in chop. Mean reversion does the opposite. Quality factors look brilliant on a ten-year arc and look catatonic during the two-year windows in the middle of it. The question every researcher eventually asks isn't which model is best. It's which model is best right now, and how do I know when right now ends.
Shishin's answer is to stop pretending one model fits every regime. Instead, we run four, each a fully independent long-bias US-equity strategy tuned for a specific market state. A breadth-driven macro classifier picks which engine fires each day. They never overlap; on any given session at most one engine is in the trade path.
The four guardians
Genbu · 玄武 · The tortoise of the North
Genbu is the quiet engine. It looks for durable margins, conservative leverage, and the absence of disasters, small-cap quality names that compound through neglect. Over the five-year window it was the active guardian only 72 days out of 1,258 , it sits out the great majority of sessions. Across those windows it took just 42 trades and contributed 12.5% of the stack's gains. The restraint is the point: Genbu doesn't fire often, but when the macro is signalling look for quality, not story, it earns its slot, and the rest of the time it stays out of the way.
Suzaku · 朱雀 · The phoenix of the South
Suzaku is the workhorse. It hunts small-cap momentum breakouts, compression-to-expansion patterns, base-and-go setups, the thrust that comes off a tight twenty-day range. It was the active guardian 383 days, more days than any other engine, taking 132 trades and producing the bulk of the stack's return: 70.4% of total gains. It is also where the stack's risk would concentrate if left unchecked, which is exactly why the macro switch rotates out of Suzaku into more defensive engines whenever breadth deteriorates.
Byakko · 白虎 · The white tiger of the West
Byakko is the bear-market translator. When the broad equity market is weak the conventional advice is “go to cash.” Byakko's observation is that there is usually a sector that isn't, commodity producers, energy explorers, gold miners. The engine trades a small basket of defensive-sector ETFs. It was the active guardian 304 days, most of them sessions on which a long-only momentum engine would have been bleeding, taking 155 trades for a 6.8% contribution. The function is regime translation: capital stays in equities, but in the equities that work when the broad market doesn't.
Seiryū · 青龍 · The dragon of the East
Seiryū is the recovery engine. Markets transition messily out of bear regimes, false starts, head fakes, a few weeks of large-cap leadership that doesn't broaden. Seiryū hunts speculative large-cap recoveries in those transition windows, when small-cap momentum hasn't earned the right to fire but cash is the wrong answer. Active just 42 days across the window, the fewest of the four, it took 24 trades for an outsized 10.3% contribution: the highest gain per day-on-duty of any engine. Most regimes don't need it; the macro deploys it surgically when the read is the bottom is in, but it's the big names leading.
The macro switch
Engines don't decide which engine is on. The classifier does. It reads the universe daily, the percentage of names above their fifty-day moving average with a fifty-over-two-hundred stack, the percentage in confirmed downtrends, and a handful of derived breadth metrics, and assigns each session a regime. The regime chooses the engine. The engine produces signals. The signals are sized by composite score and filled market-on-close, in the same closing auction the daily-bar backtest fills at.
It is deliberately a daily decision. We don't want a regime flag that flickers on a single bar; we don't want one that confirms a regime change three weeks after the market did. The seven-regime taxonomy is small enough to be intelligible, large enough to capture the meaningful states, and the persistence rules around regime transitions are tuned to favour the regimes the system has historically handled well.
What this looks like in practice
Over the 1,258-day backtest, the four-engine stack produced a 137.2% CAGR with a Sharpe ratio of 2.44 and a maximum drawdown of 15.9%. On 457 of those days , roughly 36% of the window, the system was simply in cash. No regime called for a trade, no engine fired, the capital sat. The edge isn't being always-on. The edge is being on for the right reason.
The capital-efficiency layer
The four engines decide what to hold; a second question is how hard the capital works. One answer ships in the live config: a fixed position cap, a four-position floor, that keeps the book concentrated in its best ideas without letting any one run to a reckless weight. Two other answers were tested and set aside: a conviction-funded rebalance of the stock book, and a gated top-up into confirmed winners. The top-up, even hard-gated, added roughly nothing on the locked base, so it was not adopted, its own note walks through why. The one capital-efficiency idea that advanced is an idle-cash ETF rotation sleeve, a momentum-ranked, self-hedging basket for the uninvested balance, currently in paper-observation, not yet live-trading and reported separately, outside the headline track record.
Why four, not one
A natural objection is that four engines is just an ensemble in disguise: blend them, share the capital, take a weighted average of the signals. We tried it. It under-performs. The reason is mechanical: the engines have opposing exposures to regime variables (Suzaku is risk-on, Byakko is risk-off), so blending them produces a smoothed return that doesn't capture the asymmetric upside of being correctly positioned in any single state. These sleeves are not uncorrelated, they are regime-opposed, so holding them at once averages risk-on and risk-off into a muted middle. Commitment to the regime's engine, conditional on signal, compounds faster than the blend.
The bet is that the macro classifier, a single decision per day, daily-bar resolution, deterministic logic, is good enough to put the right engine in front of the right week. A five-year backtest suggests it is, the live, paper-traded track (begun 2026-06-08) is young and is reported separately. The coming articles in this series unpack the pieces: how the composite score that ranks each engine's candidates is built, how the macro classifier reads the tape, how drawdown is treated as a metric rather than a side-effect, and what happens when you publish every trade in public from day one.