Small-Cap Value With No Dead Companies

1. Survivorship / delisting bias (optimistic). Excluding delisted names removes the companies that went to zero or were acquired at distressed prices. In small caps, typically a few percent of the universe disappears each year, with involuntary delistings realising large negative returns. The literature puts the effect at roughly 1–4 percentage points of annual return for small-cap universes - larger where the strategy tilts to distressed value names, which is exactly this strategy.

2. Index-membership look-ahead (optimistic). Applying today's Russell 2000 Value membership backwards means the model only ever considers firms that survived and remained value enough to be index members in 2025. This conditions on the outcome and is not merely survivorship - it also embeds the index provider's own rebalancing decisions, which were made partly on realised performance.

3. Restated fundamentals (optimistic). Using a current snapshot gives the model revised EPS and book values. Revisions correlate with subsequent performance, so the feature contains information that did not exist at the decision date. Order of 1–2 percentage points annually in a value strategy.

Can they explain 9.1pp? Plausibly a majority of it. Add a realistic cost estimate and the residual alpha is small and of unknown sign: at 25bp the cost drag is 0.45%/yr; at a realistic 120bp one-way in these names it is 180% × 1.20% ≈ 2.2%/yr. Combining 2–4pp survivorship, 1–2pp restatement, part of the membership effect and 1.5–2pp of extra cost, a genuine 2–4% gross edge is entirely consistent with the reported 9.1pp. But the honest answer is that you cannot sign the residual by subtraction - the biases interact and are not additive constants.

The 3% haircut is not conservative - it is arbitrary. Problems:

• The biases are known in direction but not in magnitude for this specific universe and period; a fixed haircut embeds a guess as if it were a measurement.
• The drags are not constants: survivorship bias is a function of the delisting rate in each sub-period, and revision bias is a function of how much accounting restatement occurred. Both vary through time and compound.
• Applying a haircut to a compounded CAGR is not equivalent to removing a per-period bias, and it leaves the path (drawdowns, volatility, Sharpe) uncorrected - so every risk statistic in the deck remains wrong even after the haircut.
• It launders an unverifiable number into a presentable one. If the adjusted alpha is the input to a capacity decision, the adjustment must be reproducible.

Capacity check. Total daily liquidity across 150 names at $700k each is $105m. At a 10% participation cap that is $10.5m of trading per day. Annual one-way turnover at $1.5bn is 1.8 × $1.5bn = $2.7bn, which requires 2.7bn / 10.5m ≈ 257 trading days - more than a full year, before any growth. Per name, a 1/150 position is $10m against $700k ADV: 14 days of volume to exit, or 143 days at a 10% cap.

The strategy is therefore not investable at $1.5bn on its own stated turnover. Capacity is not a single number to be asserted; it is the AUM at which turnover can be executed inside the participation constraint, which here is roughly an order of magnitude smaller.

Data and process rebuild (minimum acceptable standard):

1. Point-in-time universe constructed monthly from a survivorship-bias-free security master, with index membership history replayed as of each formation date, not read from today's file.
2. Delisting returns included, so every company that disappeared does so with its realised loss (or acquisition premium), not as a silent exclusion.
3. Point-in-time fundamentals from a snapshot table keyed to the filing date, with a realistic reporting lag, so no restated figure is visible before it was published.
4. Cost model as a function of participation rate and ADV, calibrated on the strategy's own names, replacing the flat 25bp.
5. Capacity curve: the same backtest run at $100m, $250m, $500m and $1bn, showing net return decay. If the edge is gone by $250m, the marketing claim should say so.

What I would want to see in the revised presentation:

• Gross-to-net bridge line by line (survivorship, revision, cost, capacity) rather than a single haircut.
• Sub-period results, especially 2008–09 and 2020, where delisting bias concentrates.
• Results with the worst 5% of realised delistings included, as a sensitivity band rather than a point estimate.
• An out-of-sample or held-out period the team did not use for design decisions.

Governance point. The marketing claim as drafted - '3% haircut, alpha exceeds 6%' - should be withdrawn now, regardless of the rebuild timeline, because the underlying numbers are known to be constructed with look-ahead and survivorship.