What is futuretesting?

Updated September 2026 · Written by QuantGPT, which makes futuretests, so read it knowing that

Futuretesting replays every past episode of one kind of world event against a portfolio, day by day. You pick the kind of event (an oil supply shock, a US tariff shock, a crisis in the Taiwan Strait) and the book (tickers and weights). The futuretest takes each past episode of that kind, applies what the book's holdings really did in the trading sessions after it, at today's weights, and draws the result as a cone: every episode's path, the median path, the band around it, the worst and the best, and the sample size, n. Futuretests read history; they do not forecast.

On this page: vs backtesting · the cone · the events · a worked example · run one · questions

// HOW IT DIFFERS FROM BACKTESTING

A backtest replays a rule over time. A futuretest replays event episodes against a book. The backtest asks what a rule would have done across a stretch of history; the futuretest asks what the holdings you have now went through after every past event of one kind, and then keeps score as the real days arrive.

One does not replace the other. A backtest of a rule never looks at when the tariff shocks happened, so it cannot say what they did to a book like yours. A futuretest cannot say what a rule would have done, because it runs no rule.

// HOW THE CONE IS BUILT

• EpisodesEvery past event of the kind on the desk, at most one per 21 trading sessions (the most severe of a run), so a week of daily headlines about one story counts once. Episodes the live feeds never carried, such as the 1973 oil embargo and the WHO's COVID-19 declaration, are seeded from the public record and dated to the day they were reported.
• PathsFor each episode, the book's holdings at today's weights through the 63 trading sessions after it. A holding that did not trade at the time reads its sector fund (or the S&P 500) for that episode, and the result says so.
• ConeAt every session: the median, the band (the 10th to 90th percentile once there are ten episodes, the full range below that), the worst and the best, and n.
• ClimateOptional. Keeps only the episodes that landed in a climate like today's: the 10-year Treasury yield moving the same way over six months, the same normal or inverted yield curve, and the VIX in the same third. It shows how many it kept of how many, and under five matches it uses every episode and says so.
• HedgeOptional. Your own legs, as a percent of the book, drawn beside the original median. Nothing in a futuretest suggests a trade.

// WHAT YOU CAN FUTURETEST

Eight scenarios on record, each replaying its own real episodes:

• An oil supply shock: the 1973 embargo, Iraq's invasion of Kuwait, the Abqaiq attack, chokepoint restrictions.
• War or a blockade in the Taiwan Strait: every Taiwan Strait crisis on record, plus every war onset on the desk.
• Escalation in the Gulf: wars and strikes in the Middle East, the Abqaiq attack, the 1973 embargo.
• Attacks on Red Sea shipping: the Houthi campaign on shipping, the strikes that followed, and chokepoint restrictions.
• Russia and NATO escalation: war onsets and strikes in Eastern Europe, Nord Stream, the Przewodow missile.
• A US tariff shock: every US tariff announcement on the desk, 2018 to 2025.
• A pandemic declaration: SARS, H1N1, the Wuhan lockdown, COVID-19, and the health events on the desk.
• A Fed decision: every FOMC decision on the desk.

What if, anywhere reaches past the scenarios: up to six countries anywhere on the map and one of ten kinds of event (a war starts, a strike or an escalation, either of the two, sanctions or tariffs, a chokepoint closes, a nuclear event, a ceasefire, an election, a disaster, a health emergency). It moves every past event of the kind to the countries you pick, by where each holding earns its revenue, and sets the result beside the same book as those events really went. It is the record of other places' events, transplanted: not a forecast.

// A WORKED EXAMPLE: 21 OIL SHOCKS

Our oil-shock study, run September 14, 2026, asked the futuretest question of the simplest book there is: the S&P 500 alone. It found 21 completed oil shocks from 1974 to 2021. An episode starts when WTI crude rises 25% or more over 42 trading sessions (over two months before 1986, where the record is monthly), and trigger days up to 180 days apart count as one episode. For each one, the study read the S&P 500 forward from the trigger month on monthly closes, so it reports four horizons where a futuretest in the terminal reads every session. The terminal's oil supply shock scenario replays the named episodes listed above instead, so its cone will not match these numbers.

The median of the 21 was +14.7% a year on. The same record holds a fall of 37.5%, and one month after the trigger more of the 21 were down or flat (11) than up (10). That is why a futuretest draws the whole cone: the median is one line through a small sample, not a promise.

The sample also shrinks fast when you ask for more. Count only the faster shocks and n drops: on the same speed measure, 14 of the 21 reached a 35% rise, 7 reached 50%, 3 reached 75%, and 2 reached 100% (1973-74 and 2020). Two episodes are two stories, not a rate, which is why every futuretest prints its n.

Keeping score: the 2026 shock

The 2026 shock, triggered on March 3, 2026, is not one of the 21; it is the one being scored as it happens. One month after its trigger the S&P 500 was up 10.4%, above the best of the 21 at that horizon (+8.5%): outside the range. At three months it was up 14.9% and at six months up 17.3%, inside the range of the 21 both times. That is how a futuretest kept on record reads, inside the band, under it or over it, one day at a time. One episode says nothing about the next.

Source: QuantGPT's oil-shock study, run September 14, 2026 over the QuantGPT MCP. WTI crude from FRED (daily spot from 1986, monthly before, and front-month futures for the latest days), the S&P 500 price index on monthly closes. More on oil shocks: the oil shocks page.

// HOW TO RUN ONE

1. Sign in at terminal.quantgpt.co and open World events. The events desk is open on every plan, the Free plan included.
2. Under Open, press Futuretest. The link above opens it for you.
3. Pick A scenario on record, or switch to What if, anywhere and pick up to six countries and a kind of event.
4. Pick the book: your pins, a paper book, or tickers typed with their weights. Add your own hedge, or switch Climate: like today on, if you want either.
5. Press Run it and read the cone: each episode's path, the median, the band, the worst and the best, with n. The bad episodes stay on the page.
6. Keep it. Track it from today freezes the cone and follows your book from today's close. Arm it waits for a matching event, starts on that event's date and sends you a notice. Either way, My futuretests shows where the book sits against its frozen cone: inside the band, under it, or over it.

How many you can keep on record at once is set by the plan: 1 on the Free plan, 5 on QuantGPT ($49 a month) and 25 on QuantGPT Pro ($149 a month). A finished one stays on record and stops counting.

Your AI can run the same thing. With Claude or ChatGPT connected over MCP (setup for Claude and ChatGPT), ask it to call available_categories, switch on the event desk, and run futuretest.

No account yet? Lesson 4 plays without one: one war onset against every one on record, the S&P 500 after it, and the ripple from its country to the rest of the world.

How the cone itself has done, misses included, is public. The scorecard holds out every past episode against the cone of the others, and tracks the house futuretests that real events started.

// QUESTIONS

What is futuretesting?

Futuretesting replays every past episode of one kind of world event against a portfolio, day by day, at today's weights. The result is a cone: each episode's path, the median path, the band around it, the worst and the best, and the sample size, n. It reads history; it does not forecast, and a real event can land outside the cone.

How is futuretesting different from backtesting?

A backtest replays a rule over a stretch of calendar time and shows what the rule would have done. A futuretest replays event episodes against a book: it restarts the clock at the last close before each past event of one kind and follows the book's holdings for the 63 trading sessions after it. A backtest tests a rule; a futuretest tests a book against a kind of event, and it can then be kept on record and scored as the real days arrive.

What events can you futuretest?

On QuantGPT, eight scenarios on record: war or a blockade in the Taiwan Strait, escalation in the Gulf, attacks on Red Sea shipping, Russia and NATO escalation, a US tariff shock, an oil supply shock, a pandemic declaration and a Fed decision. A what-if covers ten kinds of event, among them a war, a strike, sanctions or tariffs, a chokepoint closing, a nuclear event, an election and a disaster, in up to six countries anywhere on the map.

QuantGPT is a research tool, not investment advice. Futuretests read history; they do not forecast, and a real event can land outside the cone. Past results, real or hypothetical, do not predict future results. QuantGPT never places a trade. See the disclosures; the terms govern.