astroquanta · about
Most of what we make is dead trials. That is the product.
The number nobody prints
A backtest ends with one number. A Sharpe, usually, printed under a curve that only goes up. What the number never carries is how many other curves were drawn before that one was chosen. Nobody asks, so nobody says. Ten. Forty eight. Four hundred. The reader is asked to price a numerator with no denominator, and most of the time the reader does.
Astroquanta exists to print the denominator. Every run hands back the survivors and the full count of what was tried, with the gate that killed each specification and the reason. The surviving candidate’s Sharpe comes with a second Sharpe beside it, reduced for the number of attempts. In the canonical run, 48 went in, 6 came out, and the best of them read 0.81 walk forward and 0.63 after the haircut for 48 trials. Those two figures together are a result. Either one alone is a story.
Where it came from
A horizontal hairline with five ticks labelled onboard, rebuild, sweep, log and plans. Each tick opens one log line describing that stage of how the agent was built.
Datasets mapped and backfilled overnight.
Hypotheses tested on point in time data.
It was useful on the first night.
The log turned out to be the product.
The denominator ships with every run. Start with a pod.
The problem we set out on wasn’t statistics. It was time. At the systematic funds we knew, a researcher’s quarter went to onboarding a vendor’s dataset, mapping its entities, backfilling its history, arguing with its licence, and correcting it so every figure reflected only what was public on each past date. The hypothesis got the time that was left.
So the first version of the agent did the plumbing. Give it a dataset and it mapped, backfilled and flagged licence terms while the desk slept. Give it a hypothesis and it rebuilt the data point in time and ran the sweep. It was useful on the first night. What turned it into a company came later, when we noticed that the log it wrote to explain itself was worth more than the survivors it returned. The log held the count.
Why the denominator matters, with numbers
The research isn’t ambiguous. McLean and Pontiff, in the Journal of Finance in 2016, tracked 97 published predictors and found their portfolio returns 26 percent lower out of sample and 58 percent lower after publication. Harvey, Liu and Zhu, in the Review of Financial Studies the same year, catalogued 316 published factors and argued that a new one should clear a t statistic of 3.0 rather than 2.0, because hundreds of tests had already been run against the same history. Falck, Rej and Thesmar at Capital Fund Management found that across 72 published strategies the median post publication Sharpe was 55 percent of the in sample figure. Bailey, Borwein, Lopez de Prado and Zhu showed in the Notices of the American Mathematical Society that with only 10 random configurations the expected best in sample Sharpe is 1.57 while every one of them has an expected out of sample Sharpe of zero.
Read together, those papers say one thing: a number found by searching is inflated by the size of the search, and the honest correction is to report the size of the search. The point isn’t that nothing works. Jensen, Kelly and Pedersen found in 2023 that most of 153 factors replicate. The point is that the count gets reported.
How the agent works
You hand it a mechanism, a reason it should persist, a universe and a horizon. It writes the specifications and runs each through four gates in a fixed order.
Point in time first, so no specification can know what it could not have known.
Walk forward second. When it kills nothing, the log says so.
Costs third, charged on every fill before any curve is drawn.
Capacity fourth, in dollars, before allocation.
Then the count. Every survivor carries the number of trials behind it, and its Sharpe is cut for that number.
What it refuses to do
It does not manage money. It does not hold client assets. It does not execute orders, route them or size them. It does not recommend any security, strategy or account. It reports what survived a fixed set of tests and how many did not, and labels every figure as modelled output. Every investment decision belongs to the customer and the customer’s own investment professionals.
It also refuses to hide the gate that killed nothing, the 42 grey cells behind the 6 green ones, or the licence flags a vendor’s contract raised. Grey is most of the screen on purpose.
Who builds it
The team is described by function because the functions are what you are buying. Research engineers who have onboarded vendor data at systematic funds and know where restatements hide. Quantitative engineers who wrote the cost model and the capacity estimate. Platform engineers who keep a run that starts at 04:12 finished by 08:00. A support desk that reads support@astroquanta.com and answers in plain sentences. Fund customers get one named research engineer.
Read the research
The sources are short and mostly public. McLean and Pontiff, “Does Academic Research Destroy Stock Return Predictability?”, Journal of Finance, 2016, for the 26 and 58 percent. Harvey, Liu and Zhu, “... and the Cross Section of Expected Returns”, Review of Financial Studies, 2016, for the 316 factors and the 3.0 hurdle. Bailey and Lopez de Prado, “The Deflated Sharpe Ratio”, Journal of Portfolio Management, 2014, for a correction that accounts for trials, their variance, track length, skew and kurtosis. Their figure is a probability, which is why our 0.63 is labelled a haircut and never a deflated Sharpe. Falck, Rej and Thesmar, “Why and how systematic strategies decay”, Capital Fund Management, 2021, for the 55 percent and the 0.85 median in sample Sharpe across 72 factors. Read them and check us against them.
Closing
Overnight. One hypothesis, 48 specifications, nobody watching.
- onboard vendor.datasetmapped 4,118 entities; backfilled 2004 to 2026; 2 licence flags raised.DONE
- rebuild point in timerestatement lag applied; index membership reconstructed.DONE
- sweep 4831 cleared walk forward; 14 cleared costs; 6 cleared capacity.6 KEPT
Illustrative output. Backtested, modelled and simulated results are hypothetical. They do not represent actual trading, they carry the benefit of hindsight, and they may differ materially from any result achieved in the market. Past performance, real or simulated, does not indicate future results.
No output of the Services is a recommendation, an offer, a solicitation, or advice about any security, strategy or account. Every investment decision is made by the customer and the customer's own investment professionals.
The denominator ships with every run. Start with a pod.
6 / 48
astroquanta