Why back-test a decision engine before trusting it?
The essentials in 30 seconds
- A back-test replays past cases with known outcomes and compares the engine’s score with what happened next.
- It checks that the equation separates failures from successes before anyone relies on it.
- EQOS Asia publishes retrospective scores of 0.085 for Theranos, 0.189 for WeWork, and 0.234 for the 2008 financial crisis, all below the 0.618 line (EQOS Asia back-test pages, as of October 2026).
- A fair back-test uses only dated, public data from the time.
- Hindsight is the main risk, so inputs and thresholds are fixed before scoring.
A model that has never met a known outcome has not earned your trust. A back-test replays the past to see what the engine would have said. This guide explains the procedure, the published EQOS Asia examples, and the questions to ask of any back-test. It is one chapter of the mathematics of a decision engine.
On this page
Compare the EQOS Asia offers and request a 30-minute pre-diagnostic.
What is a back-test of a decision engine?
A back-test applies a model to cases whose outcome is already known. The engine receives the data that existed at a past date, computes its score, and the analyst compares that score with what actually happened afterwards. The practice is common wherever models support decisions, and it answers a simple question: did the equation point in the right direction when it mattered?
The result is evidence about the equation, not about a single company. A back-test that covers several cases, several years, and several sectors tells the reader how broad the reach of the equation is. A single case tells a story, and a set of cases starts to build a record. Each case becomes a documented line in that record.
Why does a decision engine need a back-test before use?
A well-written equation can still be wrong about the world. Weights may reward the wrong factors, thresholds may sit in the wrong place, and inputs may carry less information than expected. A back-test brings these weaknesses to light while the stakes are zero, before a real capital allocation depends on the score.
It also gives users a reason to listen. An executive who sees the engine rate known failures low, using only data available at the time, has a concrete basis for trusting a new score. The back-test turns a promise into a track record that anyone can inspect, and it gives the team a shared reference for later debates.
How is a fair back-test run step by step?
Five steps keep a back-test fair. First, pick cases with a known outcome, including successes. Second, freeze the date of the analysis. Third, collect only data that existed on that date, with its source. Fourth, score the case with the equation as it stood, without adjusting weights. Fifth, compare the score with the outcome and record the result, whatever it is.
The discipline lies in the fourth step. If the analyst tunes the weights until the past fits, the test proves nothing, because the equation has learned the answer. A fixed equation and fixed thresholds, written before the test, keep the exercise honest. The figure below lists the five steps.
For the full method, with sample design, point-in-time data, hit rate and false alarm metrics, read the guide to backtesting a decision model. It adds practical detail on holding cases back for a final test, and it leads to two focused pages on look-ahead bias and overfitting.
What do the published EQOS Asia back-tests show?
EQOS Asia publishes three retrospective applications. The Theranos page uses public data from 2013 and reports a φ(S) of 0.085, with 8 of 9 dimensions in the critical zone. The WeWork page uses public data from 2017 and reports 0.189, with 7 of 9 dimensions critical. The 2008 page uses public data from 2006 and 2007 and reports 0.234, with 6 of 9 dimensions critical (EQOS Asia back-test pages, as of October 2026).
All three scores sit below the 0.618 line of the STOP zone. The pages state that the data was public at the time and that the computations are reproducible. The Codex behind these scores is described on the method page. The about page adds that the system was also validated in global pilots, from a listed supply chain group to a family office (EQOS Asia about page, as of October 2026).
Why is hindsight the main risk of a back-test?
Hindsight makes the past look more predictable than it was. An analyst who knows that a company failed may read ambiguous data as a warning, or may choose inputs that confirm the outcome. The remedy is procedural: list the inputs before looking at the result, date each one, and keep the equation fixed.
Famous failures add a second bias. A set made only of collapses cannot show how often the engine raises an alarm on a company that went on to succeed. A reader should therefore look for successes beside the failures, and should read the published retrospectives as evidence of direction, not as a forecast of future results.
Which questions should a reader ask of any back-test?
Five questions separate a solid back-test from a showcase. Were the inputs public and dated before the outcome? Were weights and thresholds fixed before scoring? Are successes included next to failures? Can a third party recompute each score? Is the method described well enough to repeat the test on new cases?
Each question has a checkable answer on the published page. The EQOS Asia pages list their sources and state that the scores come from the deterministic Codex core, so a data scientist with access under NDA can recompute them (EQOS Asia back-test and method pages, as of October 2026). Apply the same checklist to any vendor, including EQOS Asia.
How can you see a back-test applied to your own case?
A 30-minute pre-diagnostic by video, scheduled in Singapore time (SGT), applies the EQOS Asia equation to a real case, and you can ask how the published retrospectives were built (EQOS Asia contact page, as of October 2026). The call carries no commitment, and a systematic NDA protects what you share.
For a fuller study, the Study tier at US$1,369 adds the prediction pillar to the audit, and the pricing page lists every tier (EQOS Asia pricing page, as of October 2026). Bring a past decision and its outcome to the call, and you can ask how the equation reads it.
See also in this guide
Frequently asked questions
What is the difference between a back-test and a forecast?
A back-test replays a past case with the data of that date and compares the score with a known outcome. A forecast projects an unknown future. The first checks the equation, and the second uses it. The EQOS Asia Prediction pillar projects 3 trajectory scenarios from the model (EQOS Asia method page, as of October 2026).
Does a good back-test guarantee future accuracy?
A good back-test shows that the equation pointed in the right direction on past cases, and it builds a record the reader can inspect. Future cases can differ from the past, which is why the score supports a decision and the executive keeps the final call.
Which data does a fair back-test use?
Only data that existed and was public on the date of the analysis, each item with its source. The EQOS Asia Theranos page, for example, uses public data from 2013, and the WeWork page uses public data from 2017 (EQOS Asia back-test pages, as of October 2026).
How many cases make a back-test credible?
No fixed number applies, and the spread matters as much as the count. A credible set mixes sectors, years, and outcomes, successes included, and it publishes sources so each score can be recomputed. The three EQOS Asia cases are a starting point for a reader who wants to inspect the method (EQOS Asia back-test pages, as of October 2026).
Test your knowledge
1. What does a back-test compare?
- The score and the known outcome
- Two prices
- Two brands
Show the answer
The score and the known outcome.
2. Which φ(S) score does EQOS Asia publish for Theranos?
- 0.234
- 0.189
- 0.085
Show the answer
0.085.
3. Which data period does the EQOS Asia WeWork back-test use?
- Public data from 2025
- Public data from 2017
- Private data from 2019
Show the answer
Public data from 2017.
4. Where does the 2008 back-test score sit against the 0.618 line?
- Above it
- At it
- Below it
Show the answer
Below it.
5. Which main risk of a back-test does this guide name?
- Too many sources
- Hindsight
- Short delivery
Show the answer
Hindsight.
About the author
Stéphane Jambu
President, EQOS Forecast Asia
Stéphane Jambu leads EQOS Forecast Asia and works as a consultant in e-commerce, search visibility, and AI. LinkedIn profile
Sources: EQOS Asia method page (eqos-asia.com/method.html), about page (eqos-asia.com/about.html), contact page (eqos-asia.com/contact.html), pricing page (eqos-asia.com/pricing.html), and the Theranos, WeWork, and 2008 back-test pages (eqos-asia.com/case-studies.html), consulted on 21 October 2026. Page updated on 21 October 2026.