This article explains a design decision in how Shishin is built. It is educational and general, an account of why the system uses several regime-specialist engines instead of one model. It is not personalised investment advice, and nothing here is a recommendation to buy, sell, or hold any security.
The obvious way to build a stock-signal system is to train one model, tune it until the backtest looks good, and ship it. We did not do that. Shishin runs four separate engines, each a specialist in one kind of market, and deploys exactly one of them on any given day. This piece is the argument for that choice: why an ensemble of narrow specialists is more robust than a single model asked to be good at everything. It is the why. The roster and its naming, how the engines share one book, and how we grade them by regime are covered in three companion pieces; this one only makes the case for many engines over one.
The short version
Markets change character: they trend, they chop, they break down, they recover. A single model tuned for one of those regimes quietly bleeds in the others. Four regime-specialist engines, each doing one job and each deployed only when its regime is present, diversify not just returns but failure modes. Firing only the engine that fits the day beats blending all four into a muted average.
The impossible tradeoff inside one model
Ask a single model to handle every market and you have handed it a contradiction. A model tuned to ride trends buys strength and holds through noise, which is exactly right in a bull tape and exactly wrong in a bear market, where the same rules keep buying failed breakouts and hold losers into the decline. Retune it to survive the downtape and it turns defensive: quick to cut, slow to commit. Now it sits out the powerful trending stretches where most of the compounding actually happens. There is no setting that is aggressive and defensive at once. Whatever you choose, you are choosing which regime the model will be wrong in.
The usual response is to split the difference: a moderate model that is acceptable everywhere and excellent nowhere. That averages the parameters, and it also averages the performance. You get a system that gives back a chunk of every bull run and still takes real damage in every bear, because it is always half-tuned for a regime it is not in. The compromise does not remove the tradeoff. It just spreads the cost across all conditions instead of concentrating it in the ones the model was built to skip.
Specialists diversify failure, not just returns
The familiar case for diversification is about returns: hold things that do not move together and your equity curve is smoother than any one holding. The case for multiple engines is stronger and less obvious. It is about diversifying how the system fails.
A single model has a single failure mode. When the market shifts into the regime it handles worst, everything the model does is wrong at the same time, and the drawdown arrives all at once. Independent specialists fail in different conditions and for different reasons. The engine built for trending markets is at its worst precisely when the defensive engine is at its best, and the other way round. Because their weak spots do not line up, the system is never fully exposed to any one of them. That is the real payoff: not a slightly higher average return, but the removal of the single point of failure that a one-model design carries by construction. We treat that idea as an evaluation discipline too, which is why we measure results conditional on the regime rather than as one blended average that hides where each engine actually earns and loses.
Four roles for four regimes
The four engines are named for the guardians of the four directions, and each is matched to a market it is built to own. The roster and its naming belong to the four-guardians framework; here the point is only that each one is a narrow specialist, not a general-purpose model:
- Genbu, the quality engine. Built for orderly, constructive markets where leadership is durable. It favours what is holding up on quality rather than chasing the loudest move.
- Suzaku, the momentum engine. The workhorse, deployed most often. It is built for strong, trending tape and presses hardest when breadth is broad and breakouts are working.
- Byakko, the defensive engine. Built for weak and falling markets. Its job is to keep capital working in the corners that hold up when the index does not, or to step aside entirely.
- Seiryu, the recovery engine. Built for the turn: the transition out of a decline, when the tape is repairing and the first real leaders are emerging from the wreckage.
None of these would be a good general-purpose model on its own. Genbu run alone would lag a raging momentum market; Byakko run alone would earn little through the long stretches when the market simply rises. Each is deliberately good at one thing and content to be idle the rest of the time. The strength is in the roster, not in any single member.
Why gating beats always-on blending
Having four specialists raises an obvious question: why not run all of them every day and blend their signals? The answer is that blending reintroduces the exact problem the specialists were built to escape.
A daily read of market breadth decides which single regime the market is in, and Shishin deploys exactly one engine into a single shared book, or none when nothing qualifies (the idle cash then goes to an ETF rotation sleeve, unless its own off-switch has it in cash too). That is regime-gating: commit fully to the engine that fits, and give the others zero weight until their regime arrives. Blending would instead keep the defensive engine partly on during a strong bull run and the momentum engine partly on during a decline. Every day you would be diluting the engine that fits with three that do not, which is just the moderate single model rebuilt out of parts. The specialists only help if you are willing to let the wrong ones sit at zero. Conviction in the right engine, and abstention in the others, is the whole point. How the chosen engine then runs inside that one shared book, sizing, exits, and rotation, is the mechanics of one book, four engines.
Gating does put weight on the decision that selects the engine. If the regime read is wrong, the system deploys the wrong specialist, and a blend would have hedged that error. We accept the exposure deliberately. The breadth-driven read that makes the call is described in the macro-regime classifier, and its most important output is the cash day: when no regime reads cleanly, the honest move is to deploy nothing rather than force a blend.
The honest limit: more engines, more ways to overfit
The case for specialists is not a licence to keep adding them. Every engine brings its own rules and its own parameters, and every parameter is another knob you can accidentally tune to the past. Push this design far enough, a dozen engines, one for every micro-regime you can name, and you no longer have a robust system. You have a curve-fitting machine with enough degrees of freedom to explain any history and predict none of the future. Four engines is a deliberate ceiling, not a starting point. Each one has to earn its place by working in a regime the others genuinely do not, and by surviving out-of-sample testing on its own, not by improving a backtest that already looks good. The discipline is a small number of well-tested engines, tested the same hard way and pruned when they do not hold up, which is exactly the graveyard documented in the experiments that failed. The argument for many engines over one only holds while each engine is justified, and holding that line is harder than adding the next engine.