αEngine: Alpha Capture from Delegated-Manager Transaction Flow

1Motivation: we already own the signal source

Alpha capture is the practice of systematically collecting external investment views and converting them into trading signals. Marshall Wace built TOPS on sell-side recommendations more than twenty years ago; Citadel has run its sell-side Alpha League platform since 2008 and, as of June 2026, is extending the model to pay external hedge-fund managers for proprietary trade ideas within its Global Quantitative Strategies division.

Attucks' manager-of-managers structure delivers a stronger version of the same raw material at zero marginal cost. Every sub-advisor executes in custody accounts whose transaction detail flows back to us daily. Unlike surveys, letters, or 13-F filings, this feed is revealed preference: actual capital committed, timestamped to the trade date, months ahead of quarterly holdings disclosures. Each trade is a vote by a professional manager we selected for skill. Until αEngine, those votes were reconciled for operations and then discarded as an information asset.

2Data and pipeline

The study covers the full custody transaction feed: 219,821 rows spanning April 2011 through July 2026, with dense coverage from 2023 onward (roughly 24,000 trades in 2023 rising to 35,000 in 2025). After normalization, 51,107 signal-eligible trades remain across 46 canonical manager firms, 14 clients, and 73 accounts.

Securities resolve to a 150,233-name Morningstar master (identifiers, GICS classification, monthly total returns) through a waterfall — ISIN, CUSIP, CUSIP-constructed ISIN, ticker, normalized name — achieving 83% resolution of equity-scope rows. Manager identity is canonicalized from 100+ raw account-name variants to one vote per firm: trades are netted per firm across all client accounts, so a model decision replicated in five accounts counts once (replication is retained separately as evidence of a firm-wide decision).

3Signal construction

For each security over a trailing three-month window, αEngine computes a conviction score (0–100) from four deterministic components:

A consensus buy requires at least two independent net buyers; a consensus sell, at least two independent net sellers. Trend detection operates in two complementary layers covering the entire market. The first is taxonomy-free discovery: every Morningstar industry the managers touch (143 monitored, 85 currently active, spanning all 11 GICS sectors) is ranked mechanically by normalized flow, persistence, and z-score against its own history — so trends surface whether or not anyone thought to name them. The second is a curated cross-sector lens (taxonomy v2, 17 themes from capital markets and energy through housing, defense, healthcare, travel, nuclear, and quantum) that captures narratives industry codes cannot express — for example, the AI complex split into hyperscalers, semiconductors, datacenter infrastructure, power, and software so rotation within the theme is visible. A crowding watch flags multi-manager pile-ins for risk management. No random numbers are used anywhere; rebuilding the artifact reproduces every figure bit-for-bit.

As of v1.4 the engine also joins the platform's portfolio and performance tables. Holdings verification: custody accounts map to monthly portfolio snapshots (34 portfolios, current through July 2026), so "new position" means verifiably absent from the pre-window snapshot, full exits are distinguished from trims, and every signal carries position weight, three-month weight change, and active weight against the account's holdings benchmark (21 benchmarks resolved). Skill weighting: each firm's consensus vote is weighted by its realized trailing-36-month information ratio from funded-account returns, bounded between 0.6x and 1.4x, so a vote from a demonstrably skilled manager counts more than one from a manager on watch. Section 5 includes a point-in-time weighted re-test: each historical cohort uses trailing IR computed only from data available at that cohort's date, so the comparison carries no look-ahead.

4Validation design

Each month from October 2023 through May 2026 forms a cohort. Within each cohort we construct the top-15 consensus-buy and top-15 consensus-sell baskets from trailing three-month flow, then measure equal-weight forward returns at one, three, and six months against the equal-weight universe of all names the managers traded in the same window — a demanding benchmark that neutralizes the managers' shared small/mid-cap habitat. Forward returns are realized Morningstar monthly totals beginning the month after cohort formation; there is no overlap between signal and measurement windows.

5Findings

5.1 The sell-side anomaly: consensus exits are premature

*Buy-basket 6m mean is skewed by a small number of outlier cohorts; the median is −1.0%, hence we do not treat it as evidence of edge. Monthly cohorts overlap at 3m/6m horizons, so t-statistics overstate independence; hit rates are the sturdier read. Panel n = 26–31 cohorts.

The names our managers collectively sold did not merely hold their own — they beat everything else the same managers were trading, persistently and at scale. The one-month cohort series below shows the pattern is broad-based rather than driven by a single episode:

Valuation corroborates the mechanism. The current consensus-sell basket trades at a median 27x forward earnings against 21x for the buy basket and 20x for the traded universe: the names being exited are disproportionately the expensive recent winners, exactly the extreme-return salience pattern documented in the academic literature on institutional selling.

Guideline-trigger classification (v1.6–v1.7) sharpens the interpretation further. Each sell vote is tested against deterministic mandate-trigger hypotheses: market-cap graduation (the name outgrew the mandate's cap range, proxied by the 95th-percentile constituent cap of the account's holdings benchmark until each IMA is codified), departure from the benchmark universe, position-size caps in both forms found in our guidelines (absolute portfolio weight and active weight versus the benchmark), sector exposure limits in both forms (absolute sector weight and overweight versus the benchmark sector weight, computed from each portfolio's holdings snapshots), and restricted lists. The result is striking: the majority of the current top consensus sells are flagged as likely mandate-driven, dominated by cap graduation — small- and SMID-cap managers selling $15–25B names their mandates no longer permit. This reframes part of the anomaly from behavioral error to structural alpha leakage: the managers are required to abandon their winners, and nothing in the program currently captures the continuation. The exit-continuation sleeve (Section 6) is therefore not a bet against our managers' judgment; it is a mechanism for harvesting returns the mandates force them to leave behind.

The academic anchor is exact. Akepanidtaworn, Di Mascio, Imas & Schmidt, “Selling Fast and Buying Slow: Heuristics and Trading Performance of Institutional Investors” (Journal of Finance 78(6), 2023, pp. 3055–3098), studied 783 institutional SMA portfolios averaging $573 million and found clear skill in buying alongside selling decisions that underperform even random-selling counterfactuals, driven by a salience heuristic — managers disproportionately sell positions with extreme recent returns, devoting far less attention to exits than to entries. Two details matter for us: their sample is separately managed accounts run for institutional clients (our exact structure), and the selling deficit concentrates among fundamentals-oriented managers running concentrated, high-tracking-error portfolios — a description of most of the Attucks stable. Our feed independently reproduces their result, in our managers, in the 2023–2026 sample.

5.2 Consensus buys: no reliable short-horizon edge — yet

The buy basket is honest noise at one and three months (t ≈ ±0.2, hit rates 41–45%). This is consistent with the literature (buying skill in the JF study manifests against a random-buy counterfactual, not necessarily against everything else the manager trades) and with mechanics: our breadth filter requires two independent buyers, which in a small/mid-cap universe often means the idea is already partially priced.

The point-in-time skill-weighted re-test (votes weighted by each firm's trailing IR as known at each cohort date) improves the basket at every horizon — one-month spread from −0.18% to +0.03%, three-month from +0.49% to +0.78%, six-month from +8.5% to +9.3% with hit rates up three to four points — but remains short of statistical significance (t ≈ 0.35 at three months). Skill weighting helps directionally; it does not yet produce a validated buy edge. We do not recommend allocating to the consensus-buy basket as a return strategy today; it retains value as an attention list and will be re-tested as feed coverage completes.

5.3 What the flow says right now (as of July 2026)

Beyond the validated anomaly, the positioning read is immediately useful. Over the trailing three months our managers rotated hard out of Information Technology (normalized flow −6.9, by far the largest sector outflow) and into Industrials (+4.1) and Financials (+2.7). Taxonomy-free discovery makes the rotation specific: the strongest accumulation in the book is regional banks (z = +3.4 with five consecutive months of net buying), followed by medical care facilities and industrial distribution, while distribution concentrates in capital-markets names, electronic components, and asset managers. Inside the AI theme, the curated lens shows the exit is surgical, not wholesale: AI Infrastructure & Datacenter names saw 3 buys against 19 sells, semiconductors 16 against 30, while AI software was balanced. Current top consensus buys include Global Industrial, Openlane, Herc Holdings, and Cytokinetics; top consensus sells include Advanced Energy, Viavi, Sterling Infrastructure, and Qualcomm.

6Investment applications

1. Exit-continuation sleeve (research). The direct monetization of the anomaly: a paper-traded sleeve holding the consensus-sell basket, rebalanced monthly, measured against the traded-universe benchmark. Six-month horizons, equal weight, no leverage. If live paper results track the backtest for 2–3 quarters, it graduates to the internal high-conviction program.
2. Manager sell-discipline oversight. The JF authors showed a simple counterfactual (random selling) would have added material annual value; our feed lets us compute each firm's realized cost of exits and put it on the table in reviews. This is an oversight capability no consultant currently offers, and it converts the anomaly from a trading edge into a manager-improvement program — Attucks' core business.
3. Risk reinterpretation. Consensus trims on funded names should stop triggering reflexive concern in portfolio reviews; historically they signal continuation. Conversely, the crowding watch flags names where genuine multi-manager accumulation creates concentration and reversal risk.
4. Quarterly positioning intelligence. Sector rotation and theme migration read directly from daily flow, months ahead of holdings-based disclosures — input to the CIO commentary and client reporting.

7Limitations

1. Sample depth. 26–31 monthly cohorts with overlapping 3m/6m windows; stated t-statistics overstate independence. The 81% six-month hit rate is the sturdier figure, but this remains a 2.5-year sample.
2. Regime dependence. The window (Oct 2023 – May 2026) contains a powerful momentum regime. Selling winners early is most costly precisely in such regimes; the edge should be expected to compress in prolonged mean-reverting markets. Quarterly re-validation is mandatory.
3. Survivorship in the return join. Basket forward returns require the security master to carry forward returns; delisted names drop from both baskets and benchmark alike, which mitigates but does not eliminate the bias.
4. Coverage. 83% of equity-scope rows resolve; several fixed-income mandates are legitimately out of scope; ~4,800 rows use imputed prices. All exclusions and the unresolved queue are visible in the module.
5. No implementation costs. Spreads are gross of transaction costs and capacity constraints, though the small/mid-cap names involved are liquid at internal-portfolio scale.

8Roadmap

1. Skill-weighted votes — shipped in v1.4 (trailing-36m IR from funded-account returns, bounded 0.6x–1.4x). Next: point-in-time skill estimates so the historical validation can be re-run weighted without look-ahead.
2. Holdings context — shipped in v1.4 (snapshot-verified initiations, trim-vs-exit classification, position and active weights). Next: extend snapshot history back beyond December 2025 for deeper verification coverage.
3. Exit-continuation paper portfolio — stand up the sleeve in Section 6.1 with monthly deterministic rebalance records.
4. Per-manager sell-cost scorecards — the counterfactual analysis for manager reviews.
5. Production pipeline — migrate the CSV ingest to a governed Supabase table with row-level security as the feed becomes recurring, and complete the historical extraction (custodian-portal report compiled as CSVs) to close the client coverage gaps.

9Reproducibility

Every number in this paper regenerates from one deterministic command: python scripts/alpha_engine/build_signals.py (signal artifact) and python scripts/alpha_engine/research_validation.py (the variant study behind Section 5). The live module is at /observability/alpha-engine; the artifact at public/data/alpha-engine/signals.json; the audit scripts (qa_artifact.py, qa_manager_dedup.py) verify one-vote-per-firm netting and coverage on demand. No stochastic methods are used.

References

Akepanidtaworn, K., R. Di Mascio, A. Imas, and L. D. W. Schmidt (2023). “Selling Fast and Buying Slow: Heuristics and Trading Performance of Institutional Investors.” The Journal of Finance 78(6): 3055–3098. doi:10.1111/jofi.13271
Hedgeweek (June 3, 2026). “Citadel expands quant platform with new hedge-fund signal-sharing initiative.”