⌁ A one-person quantitative research lab · built in public
A quant lab of one, run in the open.
“When should an automated system trust a model — and when is it too uncertain to act?”
Quant Signal Lab is one researcher building toward autonomous asset allocation — alpha research, uncertainty quantification, live out-of-sample monitoring — and publishing the process, failures included. The experiments are real; the capital, for now, is simulated.
Know when the model
doesn’t know.
01 · thesis
Most quantitative work obsesses over a single question: what happens next? This lab is organised around a harder one — not what the model predicts, but whether it should be trusted right now. Uncertainty quantification is the lab’s center of gravity: ranking models that carry their own confidence, sequential monitors that notice when a live signal starts to decay, and allocation logic that sizes with the evidence instead of against it.
A signal without its uncertainty is a guess with good branding.
The interesting failures happen out of sample. That’s where this lab lives.
Every model is wrong somewhere. The job is knowing where.
Let evidence, not ego, decide which ideas survive.
Open notebooks
02 · current researchEverything below is experimental, running on simulated capital, and liable to be proven wrong in public. That is the format working as intended.
Autonomous alpha generation
Pipelines that generate, test, and kill their own hypotheses — with pre-registered criteria doing the deciding, not enthusiasm.
Uncertainty-aware deployment
When should an equity ranking model be allowed to act? Confidence scores, abstention thresholds, and the gating logic between a forecast and a position.
Architecture for autonomous allocation
The system around the models: regime detection, risk limits, capital routing — designed to earn trust gradually, on simulated capital.
Live out-of-sample monitoring
Public dashboards that watch experimental signals meet reality in real time.1 Out of sample is where excuses go to die — watch it live ↗.
The graveyard
Failed hypotheses and dead ends, documented in full. Most ideas die; the deaths are the curriculum.
Things the lab
has actually built
03 · selected work
Anytime-valid alpha-decay detection
Sequential e-process monitors that watch a live trading signal for decay and stay statistically valid however often you peek — false-alarm control with overlapping holding-period returns, plus delay bounds on how fast a real decay gets caught.
A theorem on what uncertainty scores can’t see
“Per-Prediction Uncertainty Is Not a Concept-Shift Detector” (SSRN ↗): a detectability bound — no aggregate of per-prediction scores beats chance under strict concept shift — plus the two-level deployment architecture it forces, building on the lab’s earlier working paper ↗.
A public out-of-sample board
Experimental books marked daily against reality, with model-health and shift-attribution panels — the place to watch progress live ↗, failures included.
A hypothesis factory
A pipeline that has generated and blind-triaged over two thousand candidate trading hypotheses, with pre-registered kill criteria and a documented graveyard of everything that died.
An agentic research lab
Much of the day-to-day research runs through AI agents under machine-enforced gates: adversarial review before any result circulates, independent recomputation of any number a decision rests on, and claims that only external review can promote.
Ranking models & reinforcement learning
Cross-sectional gradient-boosted rankers, reinforcement-learning experiments for allocation and execution, and the evaluation machinery — purged walk-forward splits, leakage tests, mutation-audited test suites — that keeps them honest.
Written down,
in public
04 · papers & write-ups
Per-Prediction Uncertainty Is Not a Concept-Shift Detector ↗
A detectability bound and two-level deployment for cross-sectional equity rankers — Theorem 1 (the bound animated in fig. 1), its four corollaries, and the deployment architecture they force.
When Alpha Breaks ↗
Two-level uncertainty for safe deployment of cross-sectional stock rankers: a walk-forward study of per-stock error signals, rank-geometry coupling, and the gate between a forecast and a position.
How Long Until You Know a Strategy Is Dead? ↗
Anytime-valid alpha-decay detection on overlapping cohort returns: calibration tables, one decision formula, and the four ways the guarantee quietly voids itself. Ships with a public, self-testing monitor.
Research Program M — findings
The mathematics of model-health monitoring: detection-delay bounds for live rank-IC channels, conformal coverage that survives distribution shift, and why popular drift alarms fire when nothing is wrong. Working notes, on request.
Can AI Perform Scientific Abduction?
A research program on abductive reasoning — whether an AI system can generate the explanatory hypothesis rather than merely test it — built around matched historical discovery pairs, with an alpha-generating hypothesis engine as the end goal. Pre-registration drafted.
How the lab operates
05 · principlesEvidence over ego.
An idea survives external criticism, or it doesn’t survive. Review is invited, not endured.
Point-in-time, or it didn’t happen.
No future data, no survivorship comfort, no quietly lowered bars. If a result can’t state what was knowable when, it isn’t a result.
Kill fast; write the obituary.
A documented dead end is an asset. An undocumented one is a trap you have set for your future self.
Simulated capital until the system earns otherwise.
The ambition is real; the money, for now, deliberately is not.
Build in public.
Harsh criticism is the cheapest alpha there is.
A lab of one (n = 1)
06 · the researcherQuant Signal Lab is Ursina Sanderink. The day-to-day: reading papers, proving and stress-testing small theorems, building trading systems and the monitoring around them, finding out precisely where an assumption breaks — and iterating.
- Designing a multi-sleeve market-neutral book and stress-testing it on simulated capital
- An event-driven earnings research program, pre-registered before any result is read
- Scaling the hypothesis engine — and the abduction question behind it: can an AI system generate the hypothesis, not just test it?
- Keeping the live board ↗ honest — decay monitors on every experimental signal
A lab, not a fund.
The long-term ambition is an automated one-person fund; today nothing is managed, sold, or advised here. Strategies are experimental, the capital is simulated, and no track record is published. If you trade on anything you read here, that is on you — and frankly, it would worry me.
Be first when something launches
The research is public and free to follow. Join the waitlist and you’ll get the lab dispatches — and when something launches, whether signals, tools, or something bigger, you’ll hear it here before anywhere else.
Fellow quant or researcher? The most valuable thing you can send is criticism. Reply to any dispatch — the harshest note wins.
- Live monitoring of simulated portfolios — out-of-sample in time, but no real money is deployed. ↩