Priors, not estimates.

The factor decomposition

Value via EV/EBIT (an unlevered earnings yield — Greenblatt/Novy-Marx territory rather than book-to-price). Quality via positive FCF, positive-and-growing diluted EPS, and net debt/EBITDA under 3 — a crude QMJ proxy. Cross-sectional momentum via standard 12-1 with the reversal-month skip. Time-series momentum via the 200-DMA and six-month relative strength versus the home index. Plus an unavoidable size tilt from the market-cap floors, and a hard exclusion of financials and REITs — defensible, since EV/EBIT is meaningless for a balance-sheet business.

Value plus momentum is the well-documented pairing — Value and Momentum Everywhere — and the reason it works is the negative correlation between the premia. The theoretical backbone is sound.

Where the implementation departs from that literature

The canonical combination blends the two as weighted signals across a broad portfolio. This system combines them as a conjunction of hard gates: cheapest 30% and 12-1 > 0 and above the 200-DMA and RS > 0, plus three quality screens. Intersecting gates is not the same operation as blending z-scores. It destroys breadth, and it discards the diversification benefit that made the pairing attractive — the cheap-but-not-yet-trending name that provides the negative correlation is never held. Whether the intersection beats the blend is an empirical question this project has deliberately not asked.

A subtler asymmetry in the ranking

Both rank inputs are percentiles over the pooled UK and US survivors, so in rank space the 50/50 weighting is exactly 50/50. But value has already been truncated by magnitude — only the cheapest 30% survive, and the residual dispersion inside that bucket is economically thin. Momentum was truncated only by sign. The percentile transform hides that: nominally 50/50, the composite is materially momentum-weighted in information terms. One consequence of pooling is also worth naming: the UK/US split of the final twelve is an uncontrolled residual of whichever market screens cheaper and stronger that month — and with it, the currency mix.

Degrees of freedom — and the trade that was actually made

Count the free parameters: the 30% value cut, the rank-25 buffer, 12 positions, 200 days, 12-1 with a 21-day skip, six-month RS, the 3.0× leverage threshold, two cap floors, two liquidity floors, a 3% spread cap, the 50/50 blend. Normally that's an overfitting conversation. Here it isn't, for an unusual reason: the parameters were never fitted. Backtests are constitutionally barred from justifying any parameter, and amendment proposals must be written down before anyone checks what they would have earned. These are priors, not estimates.

That's a real defence against overfitting, and better discipline than most shops manage. The cost is that every value is also unvalidated — nobody knows whether 30% or 12 or 200 days is anywhere near optimal, and the design forbids finding out. It has traded data-mining risk for specification risk, deliberately.

The critique to lead with: statistical power

The pre-committed rule retires the system if it trails its benchmark after 36 live months. Twelve small caps, equal weight, versus a global small-cap ETF: tracking error plausibly 10–15% annualised. Over three years the standard error on the mean is roughly TE/√3 — call it 6–9%. Testing for a 3% alpha gives a t-statistic around 0.4. The 36-month verdict is a coin flip dressed as evidence. As a behavioural pre-commitment it's excellent; as a statistical test it's empty, and the project should never claim otherwise.

Related: nominal breadth is 12 positions × 12 rebalances a year, but the bets are near-perfectly correlated in factor space — effective breadth is closer to the two or three independent factor exposures. The information-ratio ceiling is low regardless of how good the signals are.

Data and estimation

No historical point-in-time fundamentals exist for this universe on free data, so no credible backtest of this specification is even possible — consistent with backtests being barred. The live loop handles look-ahead properly: the screen date bounds the data, asserted in tests, and every cycle's raw responses are archived, building a true point-in-time dataset from Cycle 1 forward. There is no risk model, no covariance estimation, no volatility scaling — equal weight is the no-estimation-error choice. And there is no sector neutralisation: unconstrained value-plus-momentum clusters by sector, so realised returns will be dominated by unintended sector bets. Of everything here, that is the largest gap between intended and actual exposure.