DOMAIN / 04 — ARTICLES
5 PUBLISHED NOTES
Engineering notes written from work I've already done — not think-pieces. Each one draws on a system that exists in this repository, and each one says where its argument stops.
10 SHORT ENTRIES — SOURCE ATTACHED WHERE ONE EXISTS
Shorter than an article and written while working: mental models, decision rules and engineering conclusions, each tied to the system that produced it.
MENTAL MODEL
A strategy's average return across paths and its return along the single path you actually trade are different quantities. Position sizing exists to keep you on the survivable side of that difference. Risk caps are not conservatism — they are what makes the time-average exist.
MARKET THINKING
Most short-horizon market movement is noise, and noise is indistinguishable from signal at small sample sizes. The practical consequence: before asking whether an edge exists, ask whether the sample is even capable of answering. 5,867 trades is a sample; 40 is an anecdote.
SOURCE — APEX-QUANT-ENGINE BACKTEST POPULATION
DECISION SYSTEM
Model fees and slippage before asking whether a strategy is profitable, not after. A setup that cannot pay its own costs is not a marginal strategy — it is a guaranteed loss. The cost-budget filter encodes this: rejection happens at admission, before capital is allocated.
SOURCE — AUTOMATED_TRADING_OS
ENGINEERING NOTE
If a backtest cannot reproduce itself, it was never a result — it was a random draw wearing the costume of a finding. Deterministic fixtures, explicit tie-breaking, exact decimals: these are not conveniences. They are the boundary between evidence and anecdote.
SOURCE — QUANT-MARKET-DATA-REPLAY
DECISION SYSTEM
A promotion gate that can be talked past is decoration. The value of governance is exactly the inconvenience it imposes — strategies graduate on verified evidence, and the gate's refusal to bend is the feature being purchased.
SOURCE — QRSIP
RESEARCH NOTE
A backtest is a statement about one historical window, and windows have weather. The same strategy in 2021-monsoon and 2022-drought produces different organisms. Regime attribution is not a refinement — it is the difference between knowing what happened and knowing why.
RESEARCH NOTE
Lookahead bias rarely arrives as reading tomorrow's price. It arrives sideways: a resample that includes the tail of the window, a join keyed on knowledge that only existed later, a timestamp rounded the wrong way. Prevention is architectural because the leaks are architectural.
SOURCE — QUANT-MARKET-DATA-REPLAY
ENGINEERING NOTE
Choosing IEEE-754 floats for prices is a modeling decision with statistical consequences — representation error accumulates through joins and aggregates until it is indistinguishable from market microstructure. Exact decimal arithmetic in research infrastructure is cheap insurance against an invisible failure class.
SOURCE — QUANT-MARKET-DATA-REPLAY
MENTAL MODEL
In layered systems — trading pipelines, data platforms, organizations — information should flow one way. When downstream stages can override upstream constraints, the architecture has no constraints, only suggestions. The 1% risk cap works because no lower timeframe can argue with it.
SOURCE — AUTOMATED_TRADING-OS
LEARNING
The craft is not finding huge edges — those are mostly noise or luck. The craft is establishing that a small edge is real: reproducible, cost-surviving, regime-aware. A 1% real edge with honest measurement beats a 50% fantasy with flattering assumptions.
CATEGORIES
Every note currently listed was written from a system in the archive — the note and the case study cover the same work, described at different resolutions.
NEXT
INSPECT THE SYSTEMS BEHIND THE NOTES