Case 04Decision support · Quant engineering
A trading engine whose best feature is saying no.نظام دعم القرار
- Year
- 2026
- Status
- Alpha · productizing
- Discipline
- Decision support · Quant engineering
- 0
- orders placed, by design
- 35
- architecture decisions recorded
- 2
- locales, natively
01 — The problem
Discretionary traders drown in signals that cannot be audited and models that sound confident. What they lack is a system that shows its maths, records what it knew and when, and refuses when the evidence is not there.
A deterministic, refusal-first trade-proposal engine for crypto and gold/FX. It evaluates market structure, multi-venue price consensus and macro events, then proposes an auditable, falsifiable trade — or declines with a typed reason. It never places an order.
02 — What I built
- 01
A pure NumPy numeric core for indicators, swing anchors, levels, sizing, R-multiples and fees — with zero LLM involvement.
- 02
Bitemporal point-in-time storage that separates when a fact happened from when the engine observed it.
- 03
A React operator terminal: candlestick charts with entry ladders, a conviction scorecard, falsification conditions and the full sizing derivation.
- 04
An empirical calibration curve — stated conviction against realised hit rate with Wilson intervals.
- 05
Bilingual English/Arabic UI with per-locale rationale, a command palette and keyboard navigation.
03 — Architecture
How the pieces fit.
- L1Terminal
- React + TypeScript
- EN/AR with RTL
- Cmd-K palette, vim-style jumps
- L2Engine
- Deterministic NumPy core
- Typed refusal reasons
- Falsification statements
- L3Evidence
- Multi-venue reconciliation
- Macro-calendar horizons
- Evidence classes: REAL / SYNTHETIC / NOT BUILT
- L4Store
- Bitemporal facts
- SQLite → TimescaleDB path
- Alembic migrations
Requests flow top to bottom · each layer depends only on the one beneath
04 — Decisions that matter
Decisions
Refusal is a product feature
Unaligned structure, an economic release inside the horizon or a failed price consensus each produce an explicit, typed refusal instead of a weak trade.
No model near the numbers
Every level and size is a pure, testable function — language models may explain, never compute.
Radical honesty about evidence
Every result in the repository is classed as REAL, SYNTHETIC FIXTURE or NOT BUILT, so a demo can never pass as proof.
05 — Security & trust
- Strictly no order placement — the terminal action is a proposal and a notification.
- Evaluations run only on request, because they touch live venues and write to the store.
- A `doctor` command that verifies every required and optional dependency before use.
06 — Stack
- Python 3.12
- NumPy
- FastAPI
- SQLite
- TimescaleDB
- React
- TypeScript
- Docker
07 — In the product
Next case · 05
Bousla
Streaming and a verified Islamic library, for one organisation.