Today’s question

Plan for the hour

A “fund” in CRSP is not one row

Incubation: how zero skill becomes “skill”

A family seeds several funds privately, then opens only the winners to the public; the private record is backfilled (Evans 2010).

Incubated funds’ alpha: 1.7% per year with the backfilled record, 0.2% after dropping the first 36 months.

The window is a design choice

• Doshi et al. (2015): drop the first 18 months
• Back et al. (2018): use returns only after the first offer date and once the fund name appears
• Dannhauser and Pontiff (2026): drop the first 3 years — also removes the burst of sentiment-driven ETF launches (Ben-David et al. 2023)

Why classification is not a detail

Dannhauser and Pontiff (2026), domestic-equity fund assets, 2005 → 2022:

Where the recipes disagree

In the handout’s simulation (where the true type of every fund is known), the recipes agree on core broad active funds and differ at the edges:

• Vendor codes (Doshi et al. 2015) let in hedged, short, and option-income funds and some target-date funds
• CRSP code + index flag (Back et al. 2018) lets in index funds whose CRSP index flag is missing (it starts late)
• Domestic-equity universe (Dannhauser and Pontiff 2026) adds sector funds by design and strips passive products

Names are data — dirty data

SP sits inside “RSP”, ind inside “Wind”, “100” in a fund name. Whole words fix some errors, not all (“Russell 2000 Alpha”). Published papers hand-check what their rules flag; Dannhauser and Pontiff (2026) report a false positive in their own data.

Three ways to know a fund’s style

Why researchers care about drift

• Benchmarks. A “growth” fund holding value stocks earns an alpha that is a style bet (Lecture 1)
• Investors’ portfolios. A growth + value mix only diversifies if each fund stays what it is
• Agency. A manager behind in the race may change strategy to catch up (Brown et al. 1996; Chevalier and Ellison 1997) — drift is one way to do it
• Sample selection. A paper on “growth funds” studies whatever the label contains

The source paper

Chan et al. (2002) — “On Mutual Fund Investment Styles”, RFS 15(5)

Finding 1 · Funds hug the market

Holdings-weighted style scores (0 = lowest stock rank, 1 = highest), CCL Table 4:

Finding 2 · Styles persist, but not tightly

Correlation between a fund’s style rank (FF3 loadings) over the past 3 years and over the non-overlapping 3 years after, year-ends 1984–1994 (CCL Table 3):

Correlations of 0.7–0.9, yet the average absolute change in rank is 0.11–0.18 — a fund moves more than a decile, on average, every three years.

Finding 3 · Losers change style

Sort funds on past two-year return (winners = top 20%, losers = bottom 20%) and on style (top / middle / bottom 30%); measure the absolute change in B/M rank (CCL Section 5.2, Table 8).

Losing value funds drift twice as far as winning ones — and toward growth (B/M rank 0.648 → 0.572). Losing growth funds stay put (0.146 → 0.171).

Is that drift active? A quick argument

A value fund that lost money holds stocks whose prices fell → their B/M ratios rose.

Finding 4 · Style-adjusted performance

Equal-weighted alphas of size × B/M fund groups (CCL Table 6 and conclusion):

• Growth minus value: about 0.143% per month (≈ 1.7% per year) under three factors
• About 1.2% per year after adding momentum
• Concentrated in small-cap funds: 2.82% per year four-factor gap

Why loadings lag: a fund we control

A simulated fund holds 70% large growth + 30% small growth, then switches entirely to value at month 121. Rolling 36-month Sharpe weights:

The estimate crosses 50% only 17 months after the switch. Short windows react faster but are noisier — the core trade-off in every returns-based drift measure.

A referee’s view of CCL (2002)

• Survivorship. The sample is funds alive at end-1997. Losers that survived may be precisely those that changed style — drift and survival are jointly selected
• Mean reversion and a mechanical gap. Estimated extreme ranks revert, and past returns depend on true style, so losers and winners differ in how noisy their ranks are. In a simulation with no drift, value losers still move ~0.016 more than value winners (handout §6.3) — the null is not zero
• No inference on the key contrasts. Table 8 and the growth–value alpha gap are descriptive sorts
• Drift ≠ decision. Rank changes mix trading, stock migration, and estimation error; separating them needs holdings and trades

Key takeaways

• The sample is the first model. Each filter removes a documented bias; skipping incubation turns zero skill into apparent skill
• Share classes are not funds. Aggregate to WFICN before counting, aging, or averaging fees
• Three recipes, three populations. They agree on the core and disagree at the edges; report robustness to a second recipe
• Styles cluster near the market and persist imperfectly (Chan et al. 2002)
• Losers drift, especially value losers moving toward growth
• Holdings beat loadings for measuring current style; loadings lag by construction

Discussion questions

1. Survivorship: how would you redo CCL’s Table 8 with the CRSP survivor-bias-free database? What result would change your conclusion?
2. Design a test that separates active drift (trading) from passive drift (stock migration) for a value fund that lost money.
3. A fund keeps its “growth” label while its returns load like a value fund. Why might the manager leave the label unchanged? How would that bias a study of fund flows?

Coming up

Next lab: run the Lecture 3 handout end to end — the filter pipeline, the three recipes, and rolling style weights for real funds (FCNTX, VWNDX).

Questions?

Calvin J. Chiou
jjchiou@nccu.edu.tw
Office: Commerce Building 261233

See you next Thursday.