A governed attribution workflow—from source data to IC pack

The problem: analysis trapped in pipelines

A typical quarterly attribution workflow crosses five or more systems: a data warehouse, a risk engine, benchmark data, research files, spreadsheets, and PowerPoint.

Analysts re-run extracts, reconcile snapshots by hand, and rebuild the investment-committee deck from scratch. A restated benchmark file can invalidate work already in circulation—and tracing every affected figure becomes a separate project.

Many teams have tried general-purpose AI assistants. The pilots often stop at the same point: portfolio data cannot leave the institution's environment, and answers without provenance cannot go in front of an investment committee.

Working the problem

Here is how that quarter-end question plays out as one governed session, drawn from real customer conversations. Every stage below is a real platform surface: the memory it recalled, the queries it ran, the evidence it attached.

The question

Why did the European Equity Alpha portfolio underperform its benchmark this quarter?

I'll break the quarter down properly. Let me pull the fund and benchmark series, run a Brinson attribution across sectors, and check whether this is a positioning problem or a stock selection problem.

Based on real customer conversations — all data fictional.

Before touching data, the agent recalls what the desk has already validated—the conventions that normally live in experienced analysts' heads.

What the team already knew

Recalled from memory

EEA benchmark conventionMSCI Europe, net EUR · validated by M. Keller

Standard attribution queryvalidated · recalled in 14 runs

Benchmark file lagconstituents arrive T-2 near UK holidays — will flag if stale

Over a few quarters, that validated memory grows to cover:

The pull is governed end to end: read-only queries against the warehouse and benchmark feed, each call carrying its audit reference.

Governed execution

warehouse.sqlQuery Q3 holdings & daily returns2,431 rows · 142 ms

marketdata.apiFetch MSCI Europe constituents & returnsQ3 2026 · daily

kernel.pythonRun Brinson-Fachler attributionsector × selection/allocation

riskmodel.apiPull ex-ante tracking error decompositionas of Sep 26

Click a call to see what ran and why · every call audited

sqlQuarter performance summary

done

SELECT p.period_week,
       SUM(h.weight * r.total_return)            AS fund_return,
       SUM(b.weight * r.total_return)            AS bench_return
FROM   holdings h
JOIN   returns r  USING (isin, period_week)
JOIN   benchmark_weights b USING (isin, period_week)
JOIN   periods p  USING (period_week)
WHERE  p.quarter = '2026-Q3' AND h.portfolio = 'EEA'
GROUP  BY p.period_week ORDER BY p.period_week;

Metric	Fund	MSCI Europe	Active
Q3 total return	+0.9%	+3.1%	-2.2
Max relative drawdown	—	—	-2.5
Hit rate (weeks)	6 / 13	—	-0.5
Tracking error (realized)	3.1%	—	+0.3

13 weeks · aggregated weekly

With both sides of the comparison in place, the attribution takes shape—decomposed in the record and charted alongside it.

The analysis

The gap is −220 bp for the quarter, and almost all of it accrued in a six-week window through mid-August. Here's the attribution — this is a selection story, not an allocation story.

Cumulative return vs benchmark

European Equity Alpha vs MSCI Europe · Q3 2026

Net of fees, EUR. Demo data.

Active risk by source

Ex-ante tracking error contribution

The finding arrives with its evidence chain attached—every figure traceable to the query, model run, or document that produced it.

The finding — evidence attached

Key finding

Two Industrials holdings drove 71 bp of the 86 bp selection loss

Both names missed on capex guidance in the same week of August. Combined active weight is 240 bp — stock-specific risk now contributes 46% of tracking error, close to the 3.0% mandate ceiling. Allocation decisions were a net positive and need no change.

4 audited tool calls · 3 memory items · 3 checks passed · as of Sep 26, 2026 closeHow was this derived?

Saved to memory

Industrials drag drivertwo names missed capex guidance the same week — candidate, pending your validation

The session ends where the quarter actually ends: a committee-ready deck, staged for review.

The deliverable

I've compiled this into an Investment Committee deck — performance, attribution, the risk picture, and recommended actions. You can present it directly or export to PowerPoint.

PresentationQ3 Performance Review — Investment Committee

5 slides

Investment Committee · October 2026European Equity Alpha — Q3 Performance Review

PerformanceA six-week selection drawdown, then stabilization

AttributionSelection effect drove −150 bp of the gap

RiskStock-specific risk is near mandate ceiling

ActionsRecommended actions

PresentExport to PowerPoint

When source data changes

A benchmark provider restates a constituent file. In a conventional workflow, that can mean tracing numbers already circulated, reconciling spreadsheets, and rebuilding the committee deck.

In DataQube, affected results can be flagged as stale, re-run against the updated source, and staged as a new draft. The previously distributed version remains intact, while the difference is visible for review.

1{ "event": "evt-d44e02", "action": "memory.recall",2  "detail": "benchmark T-2 lag - constituent file restated",3  "outcome": "applied", "run": "weekly-ic-pack/r-2406" }4 5{ "event": "evt-d44e07", "action": "result.flag_stale",6  "detail": "3 cells derived from rev-1 constituents",7  "outcome": "re-run scheduled" }8 9{ "event": "evt-d44e19", "action": "deck.stage_draft",10  "detail": "v24 staged; v23 retained as distributed",11  "outcome": "awaiting approval" }

The outcome is not an invisible overwrite. It is a reviewable draft, a visible change history, and a clear record of what changed.

The deployment: everything inside the perimeter

DataQube deploys into the institution's own environment—on-premise, private cloud, or air-gapped. The model runs within the customer's perimeter.

A governed connection layer links approved sources behind one interface. Credentials never reach the AI model. Every read is auditable, with the evidence required to understand what was accessed, why, and what result it produced.

What changes for the investment team

A trusted path from source data to the boardroom

DataQube gives investment teams one governed record across analysis, evidence, and presentation—running entirely inside the institution's environment.