Trading systems, run like experiments.

Ternio Capital builds AI-driven trading systems. Every strategy is a falsifiable claim: written down before the data, tested out-of-sample, and retired when the evidence stops holding.

The approach Research notes

FeaturesEncoderLatent statePosition

Fig. 1 — signal path, schematic edges active 000 / 000

Signal ŝₜ+0.00

PositionFLAT

State changes000

Fig. 2 — synthetic order flow driving the network above. Illustrative simulation, not a track record.

Fig. 3 — loss surface, gradient descent

θₜ₊₁ = θₜ − η ∇θ L(θₜ)

§ 01 — Notation

Most of the work is choosing what to minimise.

L(θ) = 𝔼[ (rₜ₊₁ − ŝₜ(θ))² ] + λ‖θ‖₂² Forecast error against the next return, penalised for complexity we cannot justify.

U = 𝔼[rₚ] − ½γ·Var[rₚ] − c·𝔼|Δw| The objective the live system actually maximises: return, risk aversion, and the cost of trading.

w* = Σ⁻¹μ / γ, ‖w‖₁ ≤ W Sizing under an explicit gross cap, so no single view can dominate the book.

§ 02 — Approach

Ternio: a group of three. Three people, three disciplines, one process.

Research

Ideas come from market structure, not from parameter search. We state the mechanism in a sentence before we touch the data, and we keep the sentence next to the result.

Engineering

One codebase from data ingestion to order routing, so the backtest and the live system walk the same path. Point-in-time data, no look-ahead, reproducible runs.

Risk

Exposure caps and drawdown limits are fixed before deployment and enforced by the system rather than by judgement. Transaction costs sit inside the objective, not beside it.

§ 03 — The loop

Hypothesis A written claim about why a price series should behave a certain way.

Data Tick, book and macro series normalised into a point-in-time store.

Model Sequence models and gradient boosting, constrained by priors we can defend.

Validation Walk-forward testing with costs applied, and a written record of what failed.

Deployment Small, monitored, and reversible. The loop starts again the same week.

§ 04 — Writing

We publish the parts we can share.

Method notes, failed experiments, and reading. The edge is in the execution, so the thinking can stay open.

On the half-life of intraday reversal Microstructure 2026·04

Liquidity-driven moves decay on a measurable clock. We fit that decay per instrument and use it to set holding periods, rather than picking a horizon and hoping the signal fits it. 8 min · with code

Walk-forward validation without self-deception Method 2026·02

The checklist a model has to pass before it goes near capital: expanding windows, costs applied from the first fold, and a written record of every variant we discarded. 12 min · method note

Three people, one process Notes 2025·11

How we split research, engineering and risk between us, how disagreements get resolved by a test rather than by seniority, and what we got wrong in the first six months. 6 min · notes

We are not raising. We are building.

If you work on similar problems, or want to argue with something we published, write to us.

research@terniocapital.com