Once a year, the Russell indices are reconstituted, and funds that track them must be out of deleted names on the effective date, the last Friday of June. Their yardstick is tracking error, not price, and the date is published well ahead. Selling cheap is the correct behavior for them. The bet was to stand on the other side: buy the deleted names just after the effective date, once the forced supply is exhausted, and collect the rebound.
index rebalancing / forced flow
Russell deletion reversal
The idea
How it was tested
Pre-registered on 2026-09-06, with a contaminated origin declared up front: an earlier, unregistered backtest over 2023-2025 had been seen, so those years were excluded entirely. The decision used 2015-2021, with 2022 held out. Two event types: names dropped from the Russell 2000 altogether, and names moved down from the Russell 1000 to the Russell 2000. Entry at the next open after the effective date, exit after 20 sessions, 2×ATR stop and 2.5×ATR target, smallest names first. The universe was deliberately loose, price above $3 and $200K of daily dollar volume, because deletions are the smallest names. Controls: the same design entering 5 sessions before the effective date, and the same design on new additions. Pass required a pooled alpha t of at least 1.96 and significance in at least 4 of the 7 years.
Where it died
The first two runs never reached a verdict. The first was voided by one ticker that carried two different companies' prices on alternating days, producing a single trade of +69,254%. The second was voided because the cleaning step ran after the price filter, which hid the bad rows; the output check caught it again.
The third run was clean, and it was not close. Buying deletions after the effective date lost 4.327% per trade over 280 trades, alpha t −6.49. Positive in 1 of 7 years, significant in 0. Captured in a portfolio, t was −4.228. The edge was real and strong, in the opposite direction.
The record
Thesis
Index funds tracking the Russell indices must be out of deleted names on the reconstitution effective date, whatever the price: their KPI is tracking error, not price, and the date is published a year ahead. That selling is price-insensitive, so the dislocation should revert once the forced supply is exhausted. Buy the deletions after the effective date.
Specification
- events
- Deleted from the Russell 2000 entirely; moved down from the Russell 1000 to the Russell 2000
- entry
- first session after the effective date (last Friday of June), next open
- exit
- 20 sessions, 2×ATR stop, 2.5×ATR target
- universe
- price > $3, 20d dollar volume > $200K (deliberately low — deletions are the smallest names)
- ranking
- smallest market cap first
- controls
- same design entering 5 sessions BEFORE the effective date; same design on new additions
- caveat
- Contaminated origin, declared: an earlier unregistered backtest covering 2023-2025 had been seen. Those years were excluded entirely; the decision used 2015-2021, with 2022 held out.
Gates
Pre-registrationpass
Was the hypothesis, its parameters and its pass/fail rule written down before any data was looked at?
Market-adjusted alpha (legacy)fail
Does the strategy beat SPY over the same holding windows, assuming beta = 1?
pooled alpha t ≥ 1.96 (Bonferroni, 2 event types)
alpha t-6.49alpha / trade %-4.33trades280
Measured 2026-09-07 with the engine's alpha of the time, one day before Jensen alpha replaced the beta = 1 benchmark. Not re-derived.
Gate 8 — Capture-preserved alphafail
Does the subset of trades a real portfolio can actually take keep the alpha?
t_cap ≥ 1.65 and ratio ≥ 0.50
t captured-4.23capture ratio+0.62
Measured with the portfolio simulator later found to leak slots.
Stabilityfail
Is the alpha present in both halves of history?
significant (t ≥ 1.65) in at least 4 of the 7 reconstitution years
years significant0 / 7years positive1 / 7
Why it died
Strong edge in the opposite direction: buying Russell deletions after the effective date lost 4.327% per trade (t −6.49), and five of seven years were significantly negative. A deletion is not a random shock — it happens because the company shrank, and shrinking small caps keep shrinking. New additions (−2.614%) did better than deletions, so flow direction does not explain the difference; the quality of the name does.
What it taught
The first run produced an alpha of +491% per trade from one ticker that carried two different companies' prices on alternating days (a single trade of +69,254%). The pre-registration had Bonferroni, held-out data and controls but no plausibility check on returns, and would have issued a verdict on garbage. Every pre-registration in the lab now runs that check first and is VOID if it fails — and here it caught the bad data twice.
Verdict history
- 2026-09-06Pre-registered; 2023-2025 burned because an earlier result had been seen
- 2026-09-07Run 1 VOID: interleaved ticker data (one trade +69,254%) Void
- 2026-09-07Run 2 VOID: input cleaning ran after the price filter, which hid the bad rows; the output check caught it again Void
- 2026-09-07Run 3 valid: pooled alpha t −6.49, −4.327%/trade No alpha