We build and deploy ML-driven strategies at the intersection of regime detection, signal generation, and portfolio construction. Every method is validated out of sample before capital is deployed.
Founded in 2023 — the year of our first live systematic trade — Tevvis is a research entity built around the conviction that rigorous methodology precedes any deployment of capital. The name means 23 in Marathi.
We develop strategies grounded in regime detection and machine learning. Every approach enters a funded portfolio only after demonstrating robustness across out-of-sample periods that span market regimes the model had not seen during training. The live track record is the evidence.
Research spans signal generation, portfolio construction, robustness testing, and cross-domain ML methods. Current work extends into concept drift detection and few-shot adaptation for financial meta-labelling, drawing on methods from medical imaging and other high-noise domains.
All performance figures represent real returns from a funded portfolio. Broker statements are available on request. We are a research operation, not an investment adviser.
Working papers on strategy design, ML methods in finance, and portfolio construction under uncertainty.
Establishes that the appropriate feature space depends on the question asked. Prices answer the screening question; business structure answers the clustering question. Develops a unified framework that separates these two tasks and demonstrates superior out-of-sample performance when the distinction is respected in portfolio construction.
Documents time-variation in returns to quality-screened portfolios that the existing literature's stationary characterisation does not accommodate. Uses hidden Markov models to classify regimes and analyses return distributions conditional on state, finding that momentum and quality premia exhibit the strongest regime-dependence.
Statistical framework examining whether apparent return predictability reflects genuine timing ability or compensation for time-varying risk exposure. Proposes a stationary bootstrap methodology for assessing significance of backtested performance and constructs confidence intervals for Sharpe ratio estimates under realistic autocorrelation assumptions.
Novel portfolio construction methodology using clustering algorithms as an alternative to traditional mean-variance optimisation under estimation error. Introduces a framework that distributes capital across regime-conditional sub-strategies proportional to posterior regime probabilities, demonstrating improved drawdown control and more stable alpha generation.
Cumulative outperformance since inception, through multiple distinct macroeconomic regimes, with real capital from February 2023.
| Jan | Feb | Mar | Apr | May | Jun | Jul | Aug | Sep | Oct | Nov | Dec | Year |
|---|