Position sizing as described here is the rule Shishin applies to its own paper-traded live system. It is not a recommendation for any reader's portfolio. Shishin does not publish entry prices, stop levels, or recommended position sizes for subscribers, and this article is no exception.
Equal-weighting is a confession of ignorance. It says the operator knows the names but doesn't know which one is best, and so apportions capital evenly across them. Shishin rejects the confession.
The case for sizing by conviction
A composite score is a continuous measure. A score of 90 is not the same kind of candidate as a score of 70, even if both are above the engine's eligibility cutoff. The 90-score name has, by construction, higher base rates of producing winning outcomes in historical samples. Treating both as identical 5% positions throws away the most valuable piece of information the scoring layer produced: the relative ranking.
Sizing by conviction is therefore the natural complement to scoring. The system already paid the cost of building a composite that ranks; it should also collect the benefit of using the rank in the allocation decision. Higher-conviction names get more capital. Lower-conviction names get less. A candidate at the eligibility cutoff itself gets the smallest position the rules will tolerate, or — if the portfolio is already full — none at all.
How the mapping works
The composite-to-size mapping is piecewise, monotonic, and bounded. A score in the lowest eligible band maps to a floor position size. A score in the top band maps to a ceiling position size, the largest single-name exposure the engine will take. Between the floor and the ceiling the mapping is smooth: each one-point increment in composite corresponds to a small increment in target weight, with clamps at the ends so the system never takes a vanishingly small position (operational waste) or a recklessly concentrated one (drawdown risk).
The ceiling differs by engine. Suzaku, working with small- caps that can move 6–8% on a single news bar, runs a materially smaller ceiling per name than Byakko, working with defensive-sector ETFs that rarely move more than 2% in a session. The ceilings are calibrated such that the per-engine 1-σ daily P&L contribution from a single name at the ceiling weight is approximately matched across engines. Equal dollar at risk per name was the design target; differing weight per name is the consequence.
The top-up mechanic
A position entered at a 70 composite is not the same position once the same name is scoring 85 a week later. Shishin's top-up rule recognises this. When a held name's composite climbs through a higher band, the system purchases additional shares of that name to bring the weight up to the higher band's target. The mechanism is scaled in: a name that climbs gradually accumulates weight gradually; a name that spikes in score also spikes in position size, but the mechanism doesn't permit unbounded concentration — the engine's ceiling caps the cumulative position.
Top-ups are common in healthy markets and rare in unhealthy ones — composites climb when underlying setups confirm through follow-through, which is a regime-correlated event. In a strong tape the bot will top-up names that have been held for a few weeks and have continued to score well; in chop or weakness, the system's positions tend to either exit or sit at their initial weight. The asymmetry is useful. Capital flows toward names doing what they were bought to do.
The position cap and conviction bypass
Live capital does not concurrently hold an arbitrary number of positions. The live system caps concurrent open positions at a small number — small enough that each position is materially sized, large enough to absorb the usual distribution of stop-outs without idling capital. The cap is a structural anti-dilution mechanism. Without it the roster of held names creeps up over time, the average weight per name compresses, and the system slowly turns into an equal-weight index of whatever happens to be scoring at the moment.
The cap admits exactly one exception: a candidate scoring in the top decile of the live universe — extraordinarily high composite — bypasses the cap and is admitted as an additional position. The bypass is rare by construction (the decile cutoff is set such that only a handful of names will hit it in a month) and exists for the case where a genuinely exceptional candidate appears on a day the portfolio is already populated with merely good ones. The system would rather hold one extra concentrated bet on the exceptional candidate than turn it away.
What we don't size on
Volatility-targeting (e.g., inverse-vol weighting) was evaluated and rejected. The motivation for inverse-vol — equalising the risk contribution per position — is reasonable in a long-only multi-asset portfolio, but it works against the composite's intent in a single-asset long-bias system. Inverse-vol penalises the names that recent volatility has lifted, which are usually exactly the names the composite has rewarded. The two layers were cancelling each other; the cleaner outcome was to commit to conviction-sizing and let the engine ceilings handle per-name risk.
Kelly-style sizing was also evaluated. Kelly is information- theoretically optimal under known probabilities, which is the problem: live trading does not produce known probabilities. The Kelly fraction implied by a backtest's win rate and payoff distribution is systematically too aggressive once you include the noise around those parameters, and half-Kelly is itself a function of the backtest sample. We use the composite as a direct ranking signal, not as a probability estimate, and size linearly in the rank. It is less elegant than half-Kelly. It also loses less money when the live distribution diverges from the backtested one.
Sizing is where conviction becomes capital. The composite produces conviction. The position sizing turns it into exposure. The rest is patience.