The range your book could end up in
Two weeks ahead, for the whole book and for every name in it. A range of outcomes, not a single guess.
AAPL: the real outcome landed inside the 90% range in 17 of 18 two-week windows of 2026.
Adaptive market simulation for small funds
How much could your book lose in the next two weeks?
Prices are all we need.
Tickers and weights in. The full range of your book’s next two weeks out. No data contracts, no install. Built for funds without a risk team.
The market is alive.
01Correlations doubled in four weeks
In the March 2026 sell-off, the average correlation across our 24-stock test book went from 0.10 to 0.21, then fell back to 0.08 by May.
02Not noise
Across 2026, correlations between these stocks shifted 1.6 times more from one four-week window to the next than chance alone would produce.
03Regimes turn in weeks. Risk reviews happen monthly.
So the simulation re-measures how your stocks move together from the last four weeks of prices, every time it runs.
What it shows you.
One simulation of your whole book, read six ways. Each comes with how it did in 2026.
Your bad-fortnight loss
The loss your book should exceed in only one fortnight out of twenty.
24-stock book: real losses beat it 2 times in 18, about what you would expect.
Which names drive the risk
Each position's share of the book's risk, so you can see what a trim would actually change.
Read from the same simulated paths as the range, so the two always agree.
How your stocks move together
Re-measured from the last four weeks of prices on every run, not averaged over years.
Simulated vs real co-movement: slope 0.95 across 276 pairs, no systematic tilt.
Names the model is unsure about, and why
Gap-prone names, and names whose outcomes keep landing at the edge of their range, are flagged with the reason.
24-stock book: 18 names within noise, 4 amber, 2 red.
Each stock's own shape
Fat tails, volatility bursts and lopsided moves, learned from each stock's prices instead of assumed bell-shaped.
Size of half-hour swings: 0.99 of real on average, every name within 10%.
In development
- Beta to the S&P 500 and hedge sizing
- A stressed loss number: if correlations jump to crisis levels
- Sector-grouped trims: trim one name, see the book's risk change
- Earnings dates on the confidence flags
It doesn’t predict prices.
It tells you how wide the road is, not which way it turns. In testing, the middle of the range carried no information about direction, so we never present it as a forecast.
It grades itself in public.
Trained on data to December 2025. Tested on 18 two-week windows, 2 January to 25 September 2026. Every window scored against what actually happened.
24-stock book, 2026
- 87% to 91%
- Real outcomes inside the 90% range, per stock. One year of testing, about ±15 points of uncertainty.
- 2 in 18
- Fortnights where real losses beat the bad-fortnight estimate. About 1 expected.
- 0.95
- How closely simulated co-movement tracked real co-movement, across 276 pairs (1.00 is perfect).
- 57%
- Share of real joint crashes the simulation reproduces. Our known weakness: we underestimate how often stocks crash together.
Results held from 8 to 50 stocks
| Stocks | Inside 90% range | Co-movement | Loss breaches | Joint crashes seen |
|---|---|---|---|---|
| 8 | 90% / 89% | 0.93 | 1 of 18 | 64% |
| 16 | 88% / 91% | 0.94 | 3 of 18 | 58% |
| 24 | 87% / 91% | 0.95 | 2 of 18 | 57% |
| 50 | 87% / 94% | 0.95 | 1 of 18 | 55% |
Inside 90% range: all 18 windows / once the self-correcting width is running (from window 6). At 50 stocks the bad-fortnight estimate averaged -2.8%, and the screen flagged 7 of 50 names for gap-driven tails.
Built to be checked.
A risk process you can describe to an allocator, limits included.
Method published
What goes in, what the model learns, what it measures, and how the range corrects itself.
Read the methodologyScored against reality
Five reliability checks on every test window. We publish all five, including the one we fail: joint crashes.
See the checksLimits stated up front
- One year of out-of-sample testing so far.
- Tested on large US stocks, 8 to 50 per book.
- Prices only: it doesn’t know earnings dates or news.
- It underestimates how often stocks crash together.
How it works.
Four weeks of prices, not five years of history.
- 01
Your prices
The last four weeks of half-hour prices for every name in the book. No fundamentals, no factor data, no feeds to license.
- 02
Each stock’s behaviour, learned
A generative model trained on intraday prices draws possible paths for each stock, with its own fat tails and volatility bursts.
- 03
Moving together, measured
How the names move together is measured from the same four weeks and built into the paths. The range then corrects its own width from its track record.
Already running, asset by asset.
The Distribution Feed publishes the same kind of forward range for single assets from the production pipeline: price cones, tails, drawdown ranges and volatility.
Open the Distribution FeedTalk to the people who built it.
Tell us about your book and what you would want to see about its next two weeks.
FoundersAnshuman SaxenaPrateek Sharma
Directcontact@causaquant.com