A single trade means almost nothing. Many trades, remembered across time scales, become pressure waves. The chart shows price — the waves show the force behind it. When pressure, memory, and price align, the market has structure. When they separate, it's noise. QuantDev.ai is an open-source lab that researches, tests, and publishes exactly that.
01 / Latest from the lab
Raw ticks are noise. Memory is the bridge to a backtestable alpha.
READ NOTE
Three atoms locked. One dropped on monotonicity. The data picked the seeds.
READ NOTE
Queue-normalised size. Sweep depth. Consumed-fraction. Through-ticks. Only one is its own axis.
READ NOTE
02 / The workflow
The workflow starts with raw venue data, but the objective is alpha from the beginning. We study the book to design predictive states, test those states at high-frequency horizons, build memory over the ones that survive, and then measure whether the signal has enough information and path economics to become a trading system.
We begin with the most granular feed available: market-by-order events, tick-level trades, and depth updates. Candles compress away the pressure, imbalance, and liquidity behavior we are trying to measure.
Raw data speaks before charts do.
Before there is a model, there is a market-state hypothesis: is aggressive flow building pressure, is the book absorbing impact, is liquidity refilling or vanishing? The alpha is designed as a predictive state from the start.
An alpha is a claim about market state.
Each hypothesis is tested at HFT horizons through its raw atoms: order-flow imbalance, microprice deviation, queue pressure, absorption, replenishment. If the raw atom has no predictive structure, no downstream model will rescue it.
No raw edge, no downstream mythology.
Raw microstructure effects decay in seconds, so we build memory over the ones that survive: persistence, decay, agreement, saturation, regime interaction. Memory is how short-lived pressure becomes a signal at economically realistic horizons.
Memory turns pressure into state.
A signal has to survive two independent stacks. The information stack asks whether it predicts direction — rank IC, stability, decay. The path stack asks whether the move is actually extractable — MFE, MAE, first-touch, cost hurdle.
Information first. Extractability second. Policy last.
Only after the alpha survives research does it become a system: signal definition, gates, validation, execution assumptions, monitoring, live automation. The algorithm packages the alpha. It does not invent it.
The system is the final expression of the research.
03 / The metric stack
A market-state idea does not become alpha because it sounds sophisticated. It becomes alpha only if the data supports it. The lab separates the process into three gates: information, path economics, and policy. Each gate answers a different question — and most ideas die at the first one.
Does the feature predict direction?
Rank IC, bucket shape, daily stability, horizon profile, decay, sign-flip behavior, and sample health.
Did the future path contain extractable movement?
Terminal return, MFE, MAE, first-touch barriers, cost hurdle, time-to-pay, and day-level stability.
Can the signal become a trading system?
Execution logic, gates, sizing, cooldowns, validation, monitoring, and live deployment.
04 / The research machine
The lab uses modern AI coding agents to move faster through the research loop — writing studies, refactoring harnesses, generating diagnostics, documenting failures. The edge is not just one signal. It is the research machine around the signal.
Agents help create new atom studies, memory studies, metric passes, and diagnostics.
Agents help keep the Go research stack fast, reproducible, and easy to extend.
Failed ideas are not thrown away. They become evidence for the next pass.
The lab can test more alpha hypotheses without turning research into a hand-coded bottleneck.
05 / Founder
Live
Operator
Dylan Siegel
QuantDev.ai is run by Dylan Siegel as an open-source microstructure alpha lab. The goal is to make serious market-state research visible: the raw data assumptions, alpha hypotheses, atom studies, metric stacks, failures, memory engineering, and systems that come out the other side.
Discipline
Microstructure
Approach
Open research
Stack
Go · agents
