The rules below are how Shishin manages its own backtested and paper-traded book; they are not a recommendation for any reader’s portfolio, and Shishin does not publish the exact thresholds. Figures come from a backtest over 1.258 US trading days (2021-05-12 → 2026-05-14); the rebalancing and top-up logic is backtest-validated with live wiring in progress. Past performance does not predict future results. Research output, not personalised investment advice.
Two questions decide what a long-bias book actually holds, and neither is “what do I buy.” They are: what do I sell to make room, and what do I add to when it’s working. Rebalancing answers the first; a top-up answers the second. Most of the discipline lives in getting both right.
Rebalancing is risk control, not return-chasing
The textbook case for rebalancing is about risk, not return. Left alone, a portfolio drifts: a classic 60/40 mix that is never rebalanced becomes something like 80/20 over a long bull run, quietly taking on far more equity risk than its owner signed up for. Rebalancing pulls it back to the intended exposure. The research consensus is that how often matters less than that you do it at all — annual rebalancing captures most of the benefit, and the choice between a calendar schedule and a drift-threshold trigger is mostly a cost-and-tax trade-off, not a returns one.
Shishin rebalances, but not to a fixed allocation on a calendar. Its rebalance is opportunity-funded: when the system identifies a higher-conviction candidate than something it already holds, it trims the weaker position — valued at the current market price, not the entry price — to free capital for the better idea. The trigger is not the date on the calendar; it is the arrival of a better use of the dollar. That keeps the book continuously sorted toward its strongest current convictions instead of drifting toward whatever it happened to buy first.
Adding to winners is usually a trap
The other side is adding to positions that are working — what trend-followers call pyramiding. The textbook discipline is sound in principle: add to winners, never to losers; make each add smaller than the last, because the stop is now further away and the remaining upside is smaller; and accept that the entire point is to extend the size of the rare big winners that carry a trend-following book.
In principle. In our own testing, the naive version — “a name is up, so buy more” — did not survive. We ran it explicitly and rejected it: across the sample, winners did not reliably keep winning, and the median forward return after a simple add-to-winner trigger was negative. Unconditional pyramiding is one of the faster ways to convert a good trade into a bad one, exactly as the risk literature warns. Adding to a winner is only an edge if the winner is genuinely still in the regime and setup that made it one.
…unless you gate it hard
What survived was a heavily gated top-up. The system adds to a held position only when several conditions line up at once: the broad regime still favours the engine that bought it (the momentum workhorse, Suzaku, in trending tape), the name has confirmed with a real gain rather than merely ticked up, and the add is small — a fraction of the original position — and permitted exactly once. One confirmed add, capped, in the right regime. No averaging up a second or third time; no adding in a regime that no longer supports the trade.
The gating is the entire difference between an edge and a trap. It turns “add to winners” — which fails unconditionally — into “add once to a confirmed winner the regime still endorses,” a far narrower and far more defensible bet. It also keeps the mechanism from quietly concentrating the book: a single capped add cannot run a position to a reckless weight.
The combined effect
Trimming to fund better ideas and adding once to confirmed winners are two halves of the same discipline: keep capital flowing toward the strongest current convictions and away from the weakest. Layered onto the four-engine stack alongside a position cap and the idle-cash rotation sleeve, this capital-efficiency layer is what takes the backtested book to $10.81M over the five-year window at a Sharpe of 1.87 — a higher risk-adjusted result than the engines produce on their own.
The lesson underneath both rules is the one that shows up everywhere in this research: the intuitive version of a good idea is usually too crude to survive contact with the data, and the value is in the gating. Rebalance — but to conviction, not the calendar. Add to winners — but only once, only confirmed, only in the right regime.