Quantitative Research · Trading Systems

We build trading systems
and test them on real data.

Glenrch is a self-funded company with two founders. We wrote our own market-data capture, simulator, strategy and execution code. Before a release ships, it replays recorded market days and must produce bit-identical orders, unless the change is meant to alter trading.

8exchanges integrated
700+instruments modelled
40M+order decisions simulated per day on recorded order books
01 — Product

What we're building

A trading platform we build for our own use. It records exchange market data, and a simulator replays those recordings to test strategies before any of them trade.

What we do

We will place buy and sell orders on exchanges that other traders can trade against, then hedge or close the positions that result. We will do this with our own capital only, and our revenue will be the trading profit after fees.

The problem

Exchanges need resting buy and sell orders for others to trade against. Placing them is risky: a strategy that looks good on clean data can fail in real markets because of delays, fees and the way orders actually fill. Our simulator tests strategies on recorded markets, with those effects, before they trade.

Who benefits

Once we trade live, traders on those exchanges will have more orders to trade against. The platform itself is for our own use: we do not sell it, take client money or manage money for others.

How the platform works: exchanges, capture, data archive, simulator, strategy and execution, with a research loop Exchanges Capture Data archive Simulator Strategy + execution Research 8 integrated books + trades, 24/7 grows every day replays recorded days optimization and model training, moving to GPUs orders, with risk checks: in the simulator today, live trading next
How the platform fits together. Everything runs on one codebase that we wrote. (On a phone, scroll the diagram sideways.)
Chart of recorded market data: best bid, best ask and every trade for three instruments on three exchanges, two minutes each
Two minutes each of three instruments on three exchanges, as our capture system recorded them: best bid and ask, and every trade, sized by value. This is the data the simulator replays.

Capture

Records order books and trades from eight exchanges around the clock, for 8,000+ instruments.

Simulator

Replays recorded days through the production strategy code, order by order, with exchange latency and fees. 40M+ order decisions a day.

Strategy and execution

Decides, places and hedges orders, with risk checks on every order before it is sent. Today it runs end to end in the simulator; live trading is next.

Research tools

Parameter optimization across many days and instruments, and model training on the recorded data, moving to GPUs as the models grow. Every result is tested in the simulator before use.

Development stage

  1. Jan 2026Continuous capture of live exchange feeds starts.
  2. Q2 2026Capture runs at production quality on eight exchanges.
  3. Aug 2026Enough clean history is recorded to run optimizations in the simulator.
  4. NowPre-revenue. Optimization runs fill our own research machines; nothing trades live yet.

The platform is internal, so there is no public app or sign-up; the diagram and chart above show it working. The two founders built all of it in-house.

02 — Company

Who we are

Glenrch is self-funded. We build and run systems for systematic, market-neutral trading in digital-asset markets, and we will trade only our own capital. We take no outside money, hold no client assets and keep custody for no one. Every exchange and counterparty account goes through full KYC/AML onboarding.

We wrote everything on the critical path ourselves: market-data capture, simulation, strategy, execution and risk. Capture has recorded live exchanges since January 2026 and has run at production quality since Q2 2026.

Glenrch is the brand of Glen Research LLC, registered in New Jersey, USA.

Our own capital only

We will trade only our own money. There are no outside funds, no outside investors and no client assets.

No long-term directional bets

Positions are either hedged or short-term and driven by predictions. Every position has a limit.

Limits first

Soft and hard limits and the replay check are in place before the first live trade.

03 — Platform

One codebase, from market data to orders

Capture, simulation, strategy and execution share one codebase, and everything is tested against the data that capture records.

Market-data capture

Order books and trades from eight exchanges, recorded around the clock since January 2026 and at production quality since Q2 2026, for 8,000+ instruments. The capture servers sit close to the exchanges.

Coverage

Eight exchanges integrated for market data and simulation. We record spot and perpetual markets and use perpetuals for research only. 700+ instruments are modelled, and the engine is built for 10,000.

Simulator

Replays recorded order books order by order. It reads the same files that capture writes, so strategies are tested on the markets they will trade.

Replay check on every release

Two builds run the same recorded day. If the orders they produce differ at all, the difference has to be explained before release, so a change in trading behavior shows up before it is deployed.

Audit trail

Every order, fill and decision is recorded. Any moment in production can be replayed from the same data the simulator uses.

Models

Small, fast models make high-frequency decisions, and larger models follow conditions that change more slowly. Language models read news, exchange announcements and on-chain activity, turn them into structured signals, and take part in trading decisions alongside the other models. All of them run inside the limits under Risk & Controls.

04 — Research & Compute

Research and compute

Most of our compute goes into research. Each exchange we add brings more data to replay and more models to train. Today this runs on our own servers, and we are moving the parts that grow with the data to the cloud.

Market-data archive

Order books and trades from eight exchanges and 8,000+ instruments. These markets never close, so the archive grows every day, and each new day is one more day the simulator can replay.

Simulation sweeps

We replay recorded days through the strategy code. A sweep over instruments, dates and settings splits into independent jobs and comes in bursts: hours of heavy load, then nothing. Rented capacity suits that, so each sweep can be as large as the question needs.

Model development

Fast models and slower models are built and tested on the recorded data, and larger models will be trained on GPUs. The work grows with exchanges, instruments, history and the number of settings we test, and we are early on that curve.

Language models

Reading text sources (news, exchange announcements, on-chain activity) is a workload of its own, and it grows with each market and source we add.

LLM coding agents

We use LLM coding agents every day across the codebase. The replay check is what makes that safe: every change is reviewed and must pass replay before it ships, whoever wrote it.

Moving to the cloud

Everything above runs on our own servers today, and none of the numbers on this page are cloud figures. We are moving data storage, simulation sweeps, model training and language-model inference to the cloud. Live trading will stay on small dedicated servers near the exchanges.

05 — Risk & Controls

Limits are checked before any order is sent

Risk checks run in code at four separate layers: the strategy, the order manager, the position manager and the risk manager. A final check runs just before an order goes to an exchange. Every position has a limit.

Soft and hard limits

Every order and every position has caps. An action that would break a limit is rejected before it leaves our system.

Bounded positions

Positions are hedged, or short-term and held within strict limits. We take no long-term directional bets.

Change control

No strategy or risk change reaches production without a replay on recorded market data. A change that should not affect trading must replay bit-identically; a change that does is reviewed as a diff, on top of normal code review.

Final check

Independent risk checks run on every order right before it is sent. Nothing reaches an exchange without passing them.

06 — Team

Team

Two founders. Both previously worked at Tower Research Capital, and both write the code.

AK

Alex Kshevetskiy

Co-Founder

Portfolio manager and engineer; works on applied machine learning and low-latency trading infrastructure. Ex-Tower Research Capital, then 1.5 years building systematic strategies at another trading firm. Earlier: CTO of CUTTR (Princeton, NJ; acquired by Vericept in 2005), where he built the machine-learning spam-detection models, and founder and CEO of Tellicity, a control system for stage machinery used in a concert hall.

in Profile on LinkedIn
AJ

Alexander Joly

Co-Founder

Ex-Tower Research Capital. Quantitative trader and engineer; works on strategy research and trading systems.

in Profile on LinkedIn
07 — Contact

Contact

Write to us at the address below. The founders read it.

ak@glenrch.com