Strategy Attribution and the Capital-Reallocation Decision
Live-versus-backtest decay, factor decomposition, and the retirement gate.
§ IFrame
The arc that ran every Friday since late May built one strategy and taught it to run well. Signals came first, then execution, then the cost analysis that measured what execution gave away, then the pre-trade gate that decided whether to size into a position at all, and last the runtime watch that flattens the book when a feed goes stale. By the close of that arc a strategy lives in production with a gate in front of it and a kill-switch behind it. It runs.
A book is never one strategy. It is a dozen, and capital is finite. The question this lesson answers sits one floor above the operational loop: of the strategies that run, which deserve the next dollar, which should be left where they are, and which should be turned off. Adversarial markets answer this question for you eventually, by bleeding the dead strategy's capital away one fill at a time. The discipline is to answer it first, on the evidence, before the market collects the tuition.
The instrument is attribution. Attribution asks where a strategy's return actually came from, and the answer decides whether the return is worth paying for. A strategy whose entire return can be explained by exposure to a factor you could have bought directly, at lower cost and lower drawdown, is not earning its keep. Aldridge states the bottom line plainly: when the attribution beta is high and the alpha is low, it may be cheaper to invest in the benchmark than to run the strategy (High-Frequency Trading, Ch 6, pp 181-182). That sentence is the retirement gate in one line.
§ IIFoundations
Attribution decomposes a strategy's realized return into the part explained by exposure to common factors and the part left over. The leftover is alpha, the return the strategy produced that no factor accounts for. Everything that is not alpha is beta, the return that came from riding something you did not need this strategy to ride.
The factor model makes the decomposition concrete. A strategy's return in each period is regressed against a basket of factor returns. Each factor gets a coefficient, the beta, measuring how much of the strategy's movement tracks that factor. The regression's intercept is the alpha, and its residual is the idiosyncratic return the model could not explain. Aldridge frames this through Sharpe (1992) and Fung and Hsieh (1997): reduce the many securities a strategy might touch to a small set of global factors, regress the strategy's returns on that basket, and read the alpha and betas off the fit (Ch 6, pp 174-175).
In crypto-quant the factor basket is its own design problem. The equity world has decades of agreed factors. The crypto world has a smaller, noisier set: a market factor (a BTC or total-market return), a momentum factor, a size factor across the asset cross-section, a funding-rate or carry factor for perpetuals, and a volatility factor. Chan's treatment of factor models is the working reference for building this basket from the asset's own return history rather than borrowing the equity canon wholesale (Machine Trading, Ch 2, pp 46-47). The point of the basket is not theoretical completeness. The point is that each factor in it is something you could hold directly and cheaply, so that any return the strategy explains through that factor is return you are overpaying for.
Three numbers come out of the fit and each carries a verdict. The alpha says whether the strategy adds anything a cheap factor position does not. The betas say what the strategy is secretly long or short, which is often a surprise to the person who built it. The residual variance says how much of the strategy's behavior the model cannot see, which is itself a risk because the unexplained part is the part that can turn without warning.
§ IIIMechanism — Three Reads, One Gate
Attribution in a backtest is arithmetic. Attribution in production is a discipline, because the input is a live return stream that arrives one fill at a time and the verdict it feeds is a decision to move capital. Three mechanisms turn the arithmetic into an operational gate.
The first mechanism is the live-versus-backtest decay measurement. Every strategy is born from a backtest that showed an edge. The backtest's Sharpe is the promise; the live Sharpe is the delivery, and the gap between them is the first thing to watch. Carver names the honest expectation: live performance runs below backtest performance as a rule, because the backtest never paid the full cost of being wrong in real time (Systematic Trading, Part V, pp 243-244). Track the ratio of realized live Sharpe to backtested Sharpe on a rolling window and treat a sustained collapse as the decay signal.
The second mechanism is rolling factor re-estimation. A strategy's betas are not fixed. A market-neutral strategy can drift net-long without anyone changing a line of code, because the assets it trades shifted their factor loadings underneath it. Re-estimating the regression on a rolling window catches the drift. If the strategy's return is increasingly explained by a rising market beta, the strategy is converting from an alpha engine into a leveraged long, and it should be priced as a leveraged long, which is to say cheaply.
The third mechanism is the attribution-aware capital weight. Once each running strategy has a current alpha, a current factor exposure, and a current decay ratio, the reallocation becomes a ranking. Halls-Moore frames the weight as a function of risk-adjusted return rather than raw return, so a strategy's claim on capital scales with the alpha it produces per unit of risk it takes, net of the exposure it could have gotten for free (Successful Algorithmic Trading, Part V, pp 129-130).
§ IVWorked Example — Three Strategies, One Pass
Take a paper-mode book of three crypto-quant strategies, each running in production with the operational loop from the prior arc fully in place. Capital-Zero holds throughout: this is scoring and allocation arithmetic on a paper book, and the live-capital decision stays a Sovereign-fired, per-tranche, per-wallet action that no attribution number triggers on its own.
Strategy A — funding-rate carry on perpetuals. Backtest Sharpe 1.8; six weeks live it delivers 1.5, a decay ratio of about 0.83. The rolling fit shows its return is almost entirely explained by a carry factor and almost not at all by the market factor, with a small stable positive alpha after costs. Verdict: it is doing what it said. Its return is mostly carry, but carry is the thesis, the alpha is real if small, and the decay is mild. Keep the allocation; earn a modest increase.
Strategy B — cross-sectional momentum. Backtest Sharpe 2.1; live it delivers 0.6, a decay ratio of 0.29. Worse, the rolling fit shows its market beta climbing week over week, 0.1 at inception, now 0.7. The strategy was sold as market-neutral momentum and has quietly become a leveraged long that happens to rebalance. Its alpha after the rising beta is stripped out is no longer distinguishable from zero. Verdict: the decay ratio alone warrants a cut; the beta drift confirms it. The return it still produces is market return you could buy for a fraction of the cost and drawdown. The gate trips. The strategy winds down, and its capital is freed.
Strategy C — short-horizon mean reversion. Backtest Sharpe 1.2; live 1.1, decay ratio 0.92. Its fit shows low exposure to every factor in the basket and a large but stable residual: most of its return is genuinely idiosyncratic, and its alpha after costs is clearly positive. Verdict: the rarest and most valuable kind of strategy in the book, return no cheap factor explains, delivered with minimal decay. It receives the largest share of the capital freed from Strategy B.
The reallocation that follows is not a dramatic event. It is a weekly arithmetic pass that reads three numbers per strategy, applies one gate, and adjusts the weights. The market would have reallocated for you eventually by grinding Strategy B's capital down through a hundred losing fills. The attribution pass does it in an afternoon, on the evidence, before the losses are paid.
§ VConnection to Prior Lessons
This lesson sits directly on top of the arc the prior Fridays built. The transaction-cost analysis lesson (2026-06-02) measured what execution gives away on every fill; that measured cost turns a gross backtest return into a net live return, and the net number is the only one attribution should ever see. Attribution run on gross returns flatters every strategy and retires none.
The pre-trade risk gate lesson (2026-06-04) decided, position by position, whether to size in. Attribution decides, strategy by strategy, whether to keep the position-sizer funded at all. The two gates are the same shape at two scales: a verdict that says proceed, reduce, or refuse, applied to a single trade in the first case and to a whole strategy's capital in the second. The live position monitoring lesson (2026-06-05) built the kill-switch for the fast failure, the feed that dies mid-session. Attribution is the gate for the slow failure, the edge that erodes over weeks while every feed stays healthy and every gate passes.
§ VIConnection to Today's Dev Lesson
Today's Dev lesson opens the Ruby track and builds the attribution rollup this lesson described, in Ruby's own idiom. The factor decomposition, the per-strategy alpha and beta and decay ratio, the ranking that feeds the reallocation: those are a data-transformation pipeline over a stream of per-strategy return records, and Ruby's Enumerable and Comparable modules plus its BigDecimal arithmetic are built for that shape. The Ops lesson defines what the numbers mean and what verdict they feed. The Dev lesson shows the language computing them without losing a cent to floating-point drift or a strategy to a miscompared rank.
§ VIIClosing
A strategy that runs is not a strategy that earns. The operational loop keeps a strategy alive; attribution decides whether it deserves to be. Read three numbers per strategy each week, the decay ratio, the cost-net alpha, and the factor exposure that may have swallowed the thesis, and let one gate turn the verdict into a capital move. The strategy whose return is all explained beta is a strategy you are overpaying for, and the market will eventually charge you the difference. Charge it to yourself first, in an afternoon, on the evidence.
Examine well. Reflect on this.