EventCo Budget vs Actual & Rolling Forecast: Draft Board Pack

Assembled 2026-07-20 19:14 by orchestrator.py.

DRAFT: PENDING HUMAN SIGN-OFF

This pack was assembled automatically but is not approved for distribution. No step in this pipeline sends this document anywhere. Review this pack and output/qa_report.md in full, resolve anything flagged, and sign off below before this leaves your hands.

Narrative commentary: included below from the most recent successful generation, CARRIED OVER rather than regenerated: no live model credentials were configured for this run (see output/pipeline_log.md). The QA report still checks its figures against this run's reports. Point agents/narrative_agent.py at a Claude subscription, an API key, or an organization's internal model gateway and re-run it for a freshly generated version.

1. Variance & Root-Cause

Variance Report: EventCo Budget vs Actual

Generated by agents/variance_agent.py from data/eventco_monthly_cleaned.csv (the Data Ingestion Agent's output) and the internal business notes log data/business_notes.csv. This agent never reads data/ground_truth.md.

Method

Variance = actual − budget, computed for every line item, BU and month (28 BU/line series across 30 months; full grain in output/variance_table.csv). Direction is F (favorable) or U (unfavorable): revenue above budget is favorable, cost above budget is unfavorable.

Materiality:

• Monthly (dual threshold): |variance| ≥ 10% of budget and ≥ EUR20,000. The % leg is a standard planning tolerance; the EUR floor (~0.25% of average monthly group revenue) keeps percentage noise on small opex lines out of the report, while the % leg keeps small relative wobbles on multi-million revenue lines out.
• Monthly (absolute trigger): |variance| ≥ EUR150,000 and ≥ 7% of budget. On a large line, a miss of this size (~2% of monthly group revenue) warrants comment even below the 10% tolerance; the % floor sits just above the routine forecast-noise band of the biggest lines (±6% in a normal month).
• Sustained episodes: ≥ 2 consecutive months in the same direction, each ≥ 8% off budget, with a cumulative gap ≥ EUR40,000 and ≥ 10% of the period budget. Persistent drifts are material in aggregate even when no single month clears the monthly EUR floor. A single below-threshold month is bridged when the drift resumes in the same direction immediately after, because one month of noise masking a sustained programme doesn't end the episode. A direction flip or two quiet months in a row do end it.

Root-cause discipline: this agent never invents a cause. Every material row carries exactly one of three explanation types, and the type is always visible:

1. Corroborated by the business notes log: a dated, BU-relevant note found automatically (right BU or group-wide, dated within the evidence window of 10 days before to 20 days after the variance period, and actually about that kind of spend).
2. Analyst input: no note corroborates the variance, so it went to the FP&A analyst as a follow-up; the explanation shown is what the analyst wrote back in data/analyst_commentary.csv after investigating with the business. Labeled as manual input so it can never be mistaken for machine-found evidence.
3. No clear driver identified: no note and no analyst input yet. The item stays on the follow-up list; the number is reported without a story.

Excluded from analysis: suspected data entry errors

Same magnitude rule as the ingestion agent (actual >4x or <0.25x budget is not a plausible business event). These rows are withheld from variance analysis and materiality testing entirely, pending correction at source. Explaining them as business variances would be inventing a story about a typo.

Material variances

Sorted by absolute EUR impact. Period spans of more than one month are sustained episodes; Actual/Budget/Variance are cumulative over the period. Evidence cites note IDs from data/business_notes.csv.

All other BU/line/month combinations are within materiality tolerance and are not individually commented; the full variance grain is in output/variance_table.csv.

2. Rolling Forecast

Rolling Forecast: EventCo Jul 2026 / Aug 2026 / Sep 2026

Generated by agents/forecast_agent.py from output/variance_table.csv (the Variance & Root-Cause Agent's output: cleaned actuals/budgets plus materiality flags and evidence). This agent never reads data/ground_truth.md. Last closed month: June 2026.

Method

Forecast = seasonal base × median year-over-year growth, per BU and line item:

• Base: the same month one year earlier, taken from a normalized history (see adjustments below). The events business is strongly seasonal, so same-month-last-year carries the right shape.
• Growth: the median year-over-year ratio across the last 12 closed months, on the normalized history. The median keeps any single distorted pair from steering the factor.

Normalization rules. A forecast is only as good as the history it extrapolates:

• Suspected data entry errors (still uncorrected in the cleaned file) are replaced with budget × the line's typical achievement ratio. A fat-finger digit must never touch a trend.
• One-off material months (single-month variances, explained or not) are replaced the same way. One-offs don't recur, and carrying them into a seasonal base would re-forecast last year's accident.
• Sustained episodes are kept only if still running at the cutoff (within 2 months of 2026-06); an episode that ended earlier is normalized out, because its effect is already over or absorbed into the current baseline.

History adjustments (full audit trail)

34 month-values normalized, 0 episode month-values deliberately kept.

Forecast by BU and line item

Prior year (PY) is the normalized actual of the same month last year (the forecast base). Growth is the median YoY factor applied.

Group P&L summary (forecast)

Margin: 47.9% of revenue over the horizon.

No sustained programme was still active at the cutoff, so no episode effect is carried forward in this run. Had one been active (see the adjustment rules above), its months would appear as KEPT in the audit trail and flow into the base.

3. Business Unit One-Pagers

One page per business unit, for that BU's manager: FY2025 bridge, payroll and revenue driver splits, material variances with grounded explanations, follow-ups, and the Q3 2026 outlook. Each exists as Markdown and as a print-ready PDF:

• Brand Events: output/bu_reports/brand_events.md / output/bu_reports/brand_events.pdf
• Corporate Events: output/bu_reports/corporate_events.md / output/bu_reports/corporate_events.pdf
• Digital/Influence: output/bu_reports/digital_influence.md / output/bu_reports/digital_influence.pdf
• Government & Institutions: output/bu_reports/government_institutions.md / output/bu_reports/government_institutions.pdf

4. Executive Narrative

Executive Summary: EventCo Budget vs Actual & Rolling Forecast

Three stories dominate the variances over the 30 months under review: a EUR2.08 million cost overrun on a single client project, an unfavorable currency swing on an international contract, and a cost-savings programme that ran ahead of plan. Of the twenty variances large enough to warrant comment, four are corroborated by documented business notes, fourteen carry the FP&A analyst's own commentary after follow-up, and two remain open. The rolling forecast points to continued top-line growth into the third quarter of 2026, with the group margin holding at 47.9% over the forecast period.

Variance Highlights (documented in the business notes log)

Brand Events COGS, April to June 2025: EUR2.08 million over budget (+26.6%). The largest variance in the period, tied to the Falcon product-launch event in Riyadh. The client requested a major on-site scope expansion during the build week, and the resulting overtime crews, additional staging and expedited freight were all booked at premium rates. A change order was signed with the client, but it recovered only part of the overrun, and the margin impact has been flagged to group controlling.

Digital/Influence revenue, September 2025: roughly EUR197,000 under budget (-9.3%). This is a currency effect, not a shortfall in delivery. The NovaTech roadshow contract is invoiced in USD, and the euro strengthened sharply against the dollar during the month, translating the same contracted revenue into fewer euros. Delivered scope and client commitment were unchanged, and the associated delivery costs are euro-denominated and unaffected.

Corporate Events opex, April 2025 to February 2026: EUR81,464 under budget (-16.4%), favorable. Media buying and content creation moved in-house from July 2025, ending two external agency retainers, and the business unit controller confirmed in September that the savings were tracking ahead of plan. One caveat on the dates: the reported window is the full run of consecutive under-budget months grouped by the episode test, and it starts three months before the programme did. This is the only favorable variance in the period with a documented cause in the notes log.

Government & Institutions IT opex, November 2024: EUR24,116 over budget (+158.1%). The on-site registration and AV control systems failed during a ministry summit engagement, requiring expedited replacement hardware and vendor licenses outside the normal purchasing cycle. A one-off cost with an insurance claim filed against it, not the start of a trend.

Analyst Commentary (manual input after follow-up)

Fourteen further material variances had no corroborating note in the business log, so they went to the FP&A analyst as follow-up items. The explanations below were entered manually by the analyst after investigating with the business units, and the variance report labels each one as analyst input rather than documented evidence. The largest:

Brand Events, April 2026. COGS ran roughly EUR407,000 over budget (+15.7%) while revenue ran roughly EUR389,000 above budget (+8.3%). Per the analyst's review, both trace to the same event: a client project won in late March and delivered inside April, with external production booked at short-notice premium rates. The net margin impact is slightly negative, and change-order discipline has been flagged to the project director.

Digital/Influence revenue, January 2025: roughly EUR225,000 under budget (-14.8%). The analyst attributes this to two platform go-lives slipping into February on client-side content delays; revenue is recognized on delivery, and both projects were invoiced in February.

Corporate Events, October 2024. Revenue came in roughly EUR216,000 above budget (+10.5%) with costs roughly EUR126,000 over (+17.4%): a brand activation sold for early 2025 was pulled into October at the client's request, carrying its production costs with it.

The remaining analyst-explained items are mostly timing shifts between months and delivery-mix effects, each documented row by row in the variance report.

Items Still Open

Two material variances have no documented driver and no analyst input yet: Corporate Events COGS in December 2025 (roughly EUR71,000 over budget, +10.3%) and Digital/Influence COGS in March 2025 (EUR66,500 over budget, +11.3%). Both remain unexplained and stay on the follow-up list with the relevant BU controllers rather than being assigned a cause here.

Separately, a data entry issue was identified in Brand Events' November 2025 revenue figure and has been excluded from this analysis pending correction at source.

Rolling Forecast

For July to September 2026, the group is forecast to generate roughly EUR22.6 million in revenue against roughly EUR11.78 million in costs, for a net result of about EUR10.83 million, a 47.9% margin. Revenue growth versus the same quarter last year varies by business unit: Government & Institutions +6.4%, Brand Events +6.1%, Corporate Events +5.1%, and Digital/Influence +4.6%.

The forecast is built from each business unit's own results in the same quarter last year, adjusted for its typical year-over-year growth rate, rather than from budget. One-off events, including the Falcon cost overrun and the Government & Institutions IT incident above, are deliberately excluded from that base so they are not projected forward as if they were normal, recurring activity. The Corporate Events savings programme is treated the same way: it had already concluded before the forecast cutoff of June 2026, the last closed month, so its effect is not carried into the outlook.

No cost-saving or overrun programme was still running at the June 2026 cutoff, so this quarter's outlook reflects normal seasonal business growth with no special adjustment carried forward. The two still-open variances above are the main follow-up going into the next close.

5. QA Review

QA Report: EventCo Budget vs Actual & Rolling Forecast

Generated by agents/qa_agent.py. This agent never reads data/ground_truth.md. Every check below cross-references the pipeline's own outputs against each other, the same way a reviewer would on a real, unseen monthly close with no answer key.

Status

8 passed, 0 failed, 0 skipped.

Human sign-off required. This report, and the assembled pack it accompanies, are drafts. Nothing in this pipeline is sent to anyone automatically. A human reviewer must read this report and the pack, and explicitly approve before either is distributed.

Consistency checks

• PASS: all 20 material variance row(s) carry a legitimate explanation type: 4 evidence-cited, 14 labeled analyst input, 2 explicitly open
• PASS: all 34 variance-flagged row(s) are accounted for in the forecast normalization audit trail

Data governance checks

Confirms the project's core control principle: only the Narrative Agent may call an external AI provider, and only with already-aggregated summary data, never the raw dataset, never client-identifying transaction detail.

• PASS: no agent other than narrative_agent.py references an external AI provider
• PASS: narrative_agent.py's file reads are scoped to the two aggregated report files only

Narrative grounding checks

• PASS: All 15 monetary figure(s) in the narrative trace back to a source figure within 0.5%/EUR15,000
• PASS: All 17 percentage(s) in the narrative trace back to a source figure within 1.0 pp
• PASS: narrative preserves the honesty language (3 source rows had no driver)
• PASS: excluded data-error row is mentioned without being framed as a business event

Sign-off

This pack is not final until a human reviewer completes the line below.

• Reviewed by: ____
• Date: ____
• Approved for distribution: [ ] Yes [ ] No, changes requested (see notes)
• Notes: ____