- Home
- Case studies
Case studies
Worked stories, labelled for what they are.
Capibud has no published client case studies yet. The pages below show how the product works: an explainer of the megaproject replay, three illustrative examples with invented figures and two demo workspaces built from synthetic data. Every page says which it is. The real megaproject replay and its results are shown after sign-in.
Labelled on every page. Illustrative example — figures invented; not a client or a Capibud demo Illustrative example — figures invented; not a client or a Capibud demo
All case studies
Six stories across the four decisions
Filter by type or by decision. Illustrative examples use invented figures throughout; demo workspaces use synthetic data. Neither describes a client.
How a megaproject replay works
Point-in-time public information, stepped forward quarter by quarter with no look-ahead, shown on one invented LNG project. The real replay and its results are shown after sign-in.
Read the caseA coastal refiner: one capital book across four decisions
Expansion, ventures, construction and sustaining capex pooled on one envelope: $865M of value after hurdles against $812M, with $45M less capital. Invented figures.
Read the caseAn infrastructure investor: reallocating a programme
11 positions and $6.4bn of NAV, three limits near their edge and a reallocation routed to the right approver. Invented figures.
Read the caseA plant capex round: 64 projects, one envelope
$410M of requests against a $260M envelope. Ranking by ROI misses 7 statutory items; the optimised plan meets all 11 and routes four approval batches. Invented figures.
Read the caseA climate fund: building a portfolio around the power law
An illustrative $220M fund and 18 invented companies. The optimiser earns more expected value, with a worse tail; the committee sees both.
Read the caseA data-centre developer: what to build, then keeping phase 1 on track
Three options on a 600 MW interconnect, then six construction packages where only 42% of scenarios finish within authority.
Read the caseHow we test our forecasts
A range is only useful if it holds as often as it claims
Tests 2 and 3 are below, with their limits. Test 1, the megaproject replay, is shown after sign-in; how a replay works explains its method. Test 4 is the one we would like to run with you.
Test 2
Synthetic backtest, with misspecified generators
Partial · synthetic dataWe scored Capibud's P80 against a single-point forecast on 210 synthetic projects, against a target of 80%. Under our own generator, the P80 cost held 77% of the time and the P80 finish 84%; the single-point forecast held 29% and 44%.
Read it with the caveat. Our generator contains the risk events and reference-class bias that Capibud models, so the headline is partly by construction. That is why we also score both methods against four generators built to break those assumptions: heavy tails, independent factors, no risk events and systematic underruns.
When the truth has no risk events, or projects come in under budget, Capibud is too conservative (coverage near 99%) and the single-point forecast has the lower error.
| Measure | Capibud | Single-point |
|---|---|---|
| P80 cost held, 210 projects, our generator | 77% | 29% |
| P80 finish held, our generator | 84% | 44% |
| P80 cost held, four generators built to break our assumptions | 74–99.5% | 24–85% |
| P80 finish held, the same four generators | 75–100% | 47–77% |
Synthetic projects, not customer outcomes. Coverage near 100% is not a success: it means the range is too wide.
81% inside a band built for 80%
Of 2,801 rolling 12-month forecasts on eight FRED market series since 1995, 81% fell inside Capibud's P10–P90 band. The medians run low: outcomes fell below the P50 only 37% of the time. This tests the market-factor model, not project forecasts.
Checks, not outcomes
More than 1,100 automated tests, an API test across every workspace and decision, and a one-command verification gate. These are engineering checks on demo data, not customer outcomes, and we keep the two apart.
Test 4 · yours
Run a fourth test on your own completed projects
Give us projects whose outcomes you already know. We set Capibud's range at each project's sanction date, using only what was known then, and score it against what happened, misses included.
A Decision Sprint can include this backtest. Nothing has yet been back-tested on a customer's history; the first Decision Sprints will provide that evidence, and early partners can choose to publish the calibration result with us.
Be the first case study built on real data.
Early design partners run a paid Decision Sprint and receive preferential terms in exchange for a reference and a published calibration result.