High-dimensional variable selection

T-Rex Selector

Find the real signals and control the false ones, with finite-sample guarantees.

False discoveries plague science and engineering. In genomics, a single false discovery can wrongly tie a gene to a disease, contributing to the reproducibility crisis and derailing years of research and clinical effort (Huffman et al., 2018). In finance, hundreds of studies claim to have found factors that explain stock returns, yet most of these findings are likely false (Harvey, Liu, and Zhu, 2016).

The T-Rex Selector performs high-dimensional variable selection while controlling the rate of false discoveries at a level you choose, with rigorous finite-sample guarantees, and fast enough to run on millions of variables (Machkour, Muma, and Palomar, 2025).

The problem

The Lasso finds the truth, and drowns it.

The popular Lasso method will recover the real signals, but it hands them to you mixed with a scatter of false ones. You have no way to know which is which, and no way to tune how many false discoveries you are willing to tolerate.

One dataset. Three real signals. n = 100 observations, p = 500 candidate variables — only 3 truly matter. Lasso selects the 3 — buried under 11 false discoveries T-Rex selects the 3 — and nothing else True signal (selected) False discovery Correctly ignored
Figure 1 / Same data, two methods. The Lasso buries 3 true signals under 11 false discoveries.
The fix

Pick your false discovery rate. T-Rex respects it.

T-Rex Selector controls a user-defined target false discovery rate (FDR) with a finite-sample guarantee proven via martingale theory, not an asymptotic hope. Ask for 5% and that is the ceiling you get.

You set the false discovery dial. T-Rex keeps its achieved FDR at or below your target — with a finite-sample proof. The Lasso has no dial at all. 0% 0% 10% 10% 20% 20% 30% 30% 40% 40% 50% 50% Lasso — uncontrolled (~40% false discoveries) T-Rex — achieved FDR stays under target Target FDR you request Achieved FDR
Figure 2 / Achieved FDR tracks below the diagonal: you set the target, T-Rex holds the line.

Open source

The team

The T-Rex Selector has been actively developed since 2019 by a diverse research team.

Jasin Machkour on LinkedIn Michael Muma on LinkedIn Daniel Palomar on LinkedIn Fabian Scheidt on LinkedIn Taulant Koka on LinkedIn

The papers behind the guarantee