Research

Hypothesis

If the pipeline has an edge, the names it buys should beat the benchmark after the fact.

Method

Forward-test machinery over the realized buy list, measured against SPY and against a CAPM fit rather than against a backtest — the backtest path is survivorship-contaminated and cannot answer this.

Number

Buys returned -7.08% against SPY with a CAPM alpha of -6.45% and a beta of 1.05. Of the signal lenses, fundamentals was the only predictive one (IC +0.21, t=3.9) while technical was actively anti-predictive (IC -0.28, t=-4.84) — and the two anti-predictive lenses carried 0.61 of the blend weight.

Verdict

There is no selection alpha, and the weighting was pointed the wrong way. Everything downstream — sizing, risk, exits — was being asked to rescue a signal that did not exist.

Hypothesis

If the cheap fundamental factors carry signal on a 40-name panel, they should carry it across the whole screened pool.

Method

A forecast-free factor battery over roughly 2,589 candidates per epoch from the broad screening table, escaping the restriction-of-range that a 40-name panel imposes.

Number

The strongest single factor measured +0.23 on the narrow panel and approximately zero at breadth. The narrow result was noise, and widening the panel disproved it rather than confirming it.

Verdict

The free factors are dead at breadth. Moving this would take paid data rather than a better combination of the free kind, and the anti-predictive lens was dropped from the blend.

Hypothesis

If single factors measure near zero, a shrunk composite of them might still rank — and possibly at a horizon the fund is not trading.

Method

A shrunk composite of several cheap fundamental factors, fitted walk-forward rather than in one sweep — because a fitted sweep is how the earlier regime-tilt result turned out to be an artifact.

Number

The shrunk composite reached a rank IC of +0.075 at 60 days with sign stability 0.80 — the first result in the project to clear the standing bar — against only +0.025 at 20 days. The composite beat every single factor on its own, so the "dead zoo" reading was an artifact of looking at one factor at one wrong horizon.

Verdict

The problem is vehicle and horizon, not an empty pond. The edge is quarterly and the fund rebalances weekly, which is a mismatch that costs nothing to fix. Forward validation is still pending.

Hypothesis

Each place the language model touches a decision should be separable, and each should be measurable against what happened next.

Method

A forecast-free scorecard over the logged decisions, scoring each place a language model touches a decision independently, against what happened next.

Number

The risk veto was the only positive signal. Names it vetoed went on to return -4.51% forward alpha, and the approve-minus-veto spread was +0.034 with t=2.26 on n=6. Every other surface measured approximately zero.

Verdict

Keep the veto and keep measuring it; reduce the language model where it measures zero; keep it out of selection labelling entirely. The identity that survives measurement is risk manager, not stock picker — and n=6 is thin enough that the veto result is a lead, not a conclusion.

Hypothesis

The five-lens synthesis quorum is the most expensive layer in the pipeline, so a smaller subset of lenses might do the same work for less.

Method

A forecast-free harness that re-scores already-logged proposals through the live aggregation path, so every variant costs zero additional model calls, on a common panel.

Number

The full quorum measured -0.045 on the common-panel information coefficient, and every reduced subset measured approximately zero or negative on the matured panel. An apparently strong three-lens subset (+0.345) turned out to be an artifact of scoring it with a non-production component of the pipeline over a shifting cross-section, and was caught in two rounds of cold review.

Verdict

Keep all five, because nothing justifies cutting any of them — and note that nothing shows the five earn their cost either. Both halves of that are the finding. The harness re-runs once the composite has 60 days of forward alpha.

Hypothesis

An earlier study of this project's own exits found that never selling at a loss — keeping only the profit-protection exit — would have turned the fund's loss into a double-digit gain. If that held, the drag would be the exit policy rather than the stock selection.

Method

A day-by-day replay of every recorded fill under the constraint the question requires: a position that is not sold does not release the cash that funded the next purchase, so purchases are skipped when the money is not there. The replay drives the live exit rule itself rather than a description of it, and is checked against the realized book before any counterfactual is reported — share counts match exactly on 78 days, and the simulated exit reproduced 13 of the 15 real ones with 5 false fires.

Number

The original figure does not reproduce. Holding the losers returned +3.42% against the fund's actual -3.53% — both measured on the capital the fund started with, the same basis as this site's headline figure — but at a beta of 1.62 into a benchmark that rose 5.22%, so the CAPM alpha is -8.1% with t=-0.32 and volatility is 25.7% against the benchmark's 12.2%. The spread was market exposure, not skill. The earlier +11.8% reproduces to within a point only by also returning the cash the unsold position would have tied up, which spends the same dollar twice.

Verdict

The live switch is held: it rested on a figure that credits capital the policy would never free. Sweeping the exit rule's sensitivity across its plausible range produced positive alpha at no setting — though that varied one parameter while holding the others at their live values, so it bounds the result rather than proving it. The window also straddles a change to the fund's own exit rules in August 2026: it was up 2.5% before that change and down 6.4% after, so any single whole-window figure averages across two different policies. The per-position evidence that prompted the study is untouched, and is now being measured forward rather than in-sample.

Hypothesis

Three things were tested in one grid. That trading less often helps, since this project has repeatedly found its own turnover expensive. That the exit policy running today would change that answer, since the earlier cadence study used the exit mix the fund has since left. And that an exit should be followed by buying the best-scored name not already held, rather than leaving the proceeds in cash.

Method

Twelve arms — four rebalance frequencies crossed with three exit policies — replayed day by day over every recorded fill, then run again with the cash an exit frees put back to work two different ways: spread across what is already held, or spent on the best-scored name that is not. Exits always fire when they actually fired; only the rebalance is held back. Each selection claim is scored against a null that holds the mechanism and the count fixed and varies only the choice, so "declining some exits" is separated from "declining THESE exits", and "buying something" from "buying the best-scored thing". Two of those nulls enumerate every possible choice rather than sampling.

Number

Slowing the rebalance improved annualised alpha by 22.3 points, and almost none of it was the schedule: the slowest arm averaged 61% in cash and ended 78% in cash. Forcing that cash back into existing holdings made eleven of the twelve arms WORSE. Spending it on the best unheld name recovered more than that — nine of twelve arms turn positive — but the ranking earns none of it. Against 300 random picks from the same candidates the ranked arm sits inside its own null in all twelve arms, and REVERSING the score outright did better in ten of twelve. Keeping only the losses a published solvency screen called safe moved alpha by 0.4 points and flipped sign between frequencies.

Verdict

Nothing here is claimable: no arm reaches the bar, and the largest t anywhere in the grid is 1.54 — on a NEGATIVE result. The nulls say why nothing is measurable: at roughly forty decisions, one draw landing on a name that tripled annualises into an alpha of several hundred percent. What survives is narrow and worth keeping: holding idle cash is a measurable cost, and the earlier reading that a slower cadence adds value beyond de-risking is a statement about WHICH names were held, not about holding more of them. The conviction ranking is doing no work at this horizon, which agrees with two earlier findings — that the free factors die once measured at breadth, and that this signal ranks at a quarter rather than a week.

Hypothesis

The fund currently keeps every losing position open and closes only on profit protection. If some losers are worth keeping and others are not, then two measures it already computes — a published solvency measure and its own fundamentals score — should be able to tell those apart, and replaying the fund's own history under each rule should show which reading of them is right.

Method

Six rules for when a loss may stay, crossed with four rebalance frequencies and three ways of redeploying the cash a closed position frees: seventy-two day-by-day replays over every recorded fill. The six run from keeping every loss, through requiring either measure alone, requiring both together and accepting either one alone, down to keeping none — which is the fund's actual recorded history, so the grid contains its own control. Profit protection closes winners in every arm and no rule here can decline it. The selection claim is then scored twice against nulls that hold the mechanism and the keep COUNT fixed and vary only WHICH losses are kept, so "keeping some losses" is separated from "keeping THESE losses", and once more with each kept decision removed in turn. One limit applies to every replay on this page, including the two above: the day-by-day prices come from an outside market-data vendor rather than from the fund's own records, and a vendor can restate that history after the fact — a share split corrected months later changes what a past day's return was. Nothing in these figures records which version of that history produced them, so re-running any of these replays later may not reproduce them exactly, and a difference would not by itself indicate an error.

Number

At the cadence the fund ran, with the freed cash buying the best-scored name not already held, requiring BOTH measures to approve was the best of the six at +6.7% annualised CAPM alpha against −0.8% for the rule live today and −11.8% for what the fund actually did. None of that is a result: zero of the seventy-two cells clear the bar, the largest |t| anywhere is 1.54 and sits on a negative arm, and this arm's own t is 0.44. Nor is it the best cell in the grid — it ranks 3rd of 72 by alpha and 37th by |t|, the exact middle, and the most significant cell anywhere is a −27.2% catastrophe. The ordering the comparison shows is specific to one rebalance frequency: requiring both measures beats the better single measure by 4.8 points at the cadence the fund runs, is the worst of the three at monthly, and declines nothing at all at quarterly. Across the whole grid the best single-measure cell beats the best combined one by 0.27 points, so "both measures beat either alone" is a statement about one cadence rather than a property of the rules. What the figure shows is mostly not the rule: with the freed cash left idle the same six run −8.1, −9.5, −7.7, −0.5, +1.3 and −21.9. That gap cannot be attributed cleanly, because a rule that leaves its cash idle is deciding against a different book and so keeps a different number of losses — thirteen rather than ten for the rule being rolled out. And the lever is not reinvestment as such: spreading the same cash across names already held is WORSE than leaving it idle, in twenty-two of twenty-four cells. It is buying into a name the fund did not already hold that carries the effect, which is a point about breadth rather than about either the rule or the ranking — the ranking, measured against picking at random, carries nothing. Two separate nulls then take most of it back. The choice of WHICH losses to keep clears its count-matched null, but drop any one of its nine kept decisions and it falls to the 26.5th percentile of the same field. And the rule that spends the cash earns nothing at all: the ranked pick sits inside its own random-pick null in twenty-three of twenty-four cells, and buying the WORST-scored unheld name instead did better in twenty-two. The keep count does not explain the ordering either — the two rules that keep the same nineteen losses land 6.9 points apart.

Verdict

Shipped, with the reasoning published rather than the confidence. It is the best-supported of the three readings and the project is overriding a gate it set for itself to roll it out, behind a flag, with the cold-start interlock that a dark fundamentals lens suspends the rule rather than selling everything. Two claims this study first made were withdrawn by its own follow-up: that this was the only arm in the programme to land outside a count-matched null, and that an interior optimum exists which it sits nearest. The second was an artifact of a null that withheld profit protection from the positions it kept, and at random keeping MORE losers is monotonically better — so the count's own gradient runs opposite to the explanation that was inferred from it. Nine decisions cannot settle this; the forward instrument can, and holds no matured rows yet.

If the fund had never sold at a loss

Indexed to 100 at the capital the fund started with. In the counterfactual a position closes only on profit protection — a loss is never cut. The shaded band is what declining the same number of exits AT RANDOM produces, over 240 draws. The counterfactual lands at the top of that band rather than clear of it.

n=104 daily returns · 27 of 122 recorded sells declined · 240 null draws

If the fund had reinvested every exit instead of holding the cash

Indexed to 100 at the capital the fund started with. The dotted line is what the fund actually did. The others are counterfactuals at the cadence the fund actually ran: one closes a position only on profit protection, the other also lets a solvency screen decide which losses to take — and both spend the freed cash on the best-scored name not already held. The shaded band is picking that replacement AT RANDOM, over 300 draws.

n=104 daily returns · 19 loss exits declined · 17 replacements · 300 null draws · CAPM alpha +1.9%, t=0.09

If the fund kept only some of its losses

Six rules for when a losing position is allowed to stay, each replayed over the fund's own recorded trades at the cadence it actually ran, with the cash a closed position frees spent on the best-scored name not already held. Profit protection closes winners in every rule and no rule here can decline it. Bars are annualised CAPM alpha on one fixed window, zero drawn; rows run from keeping every loss down to keeping none.

n=104 daily returns · 0 of 72 grid cells clear the alpha-claim bar · largest |t| anywhere 1.54