One trade, from the exchange
to the record it leaves behind.
Seven stages, in this order, every time. Scroll and the diagram follows — each stage is drawn as it is described, and the page changes colour as the trade moves through it.
01
Exchange
02
Analysis
03
AI
04
Risk
05
Execution
06
Journal
07
Analytics
It starts with a connection you control
You connect the venues you already trade on with keys that cannot withdraw. Feeds arrive over websockets and are normalised into one internal representation — the same candle, the same book, whatever the venue calls it. Gaps and duplicate frames are detected and backfilled before anything downstream sees them, so no strategy silently trades a hole in its own history.
- Venues
- Binance · Bybit · OKX
- Key scope
- Trade only, never withdraw
- Feed
- Normalised · gap-checked
Prices become a description of the market
Every close, the system extracts what the chart is doing rather than what it costs: trend state on each declared timeframe, swing structure, ranges and their edges, liquidity pockets, and the volatility regime. This is the layer that lets a strategy say “only in expanding volatility” and have that mean something enforceable.
- Extracted
- Trend · structure · liquidity
- Timeframes
- As declared per strategy
- Regime
- Classified on 3 horizons
Three models score it, an arbiter decides
The structure, the regime and your open exposure become one feature vector. A structure model, a regime-conditioned momentum model and an analogue-recall model each score it independently and return an attribution rather than a verdict. The arbiter weighs them by how well calibrated each has been in this regime, applies the threshold, and writes down its reasoning in plain language.
- Models
- 3 · scored independently
- Threshold
- 72 of 100, configurable
- Output
- Decision + written rationale
Then it has to get past risk
A separate service with thirteen responsibilities and veto power. Daily budget, weekly budget, exposure ceiling, correlation load, schedule and blackout windows, venue health, position count, and the sizing itself — any one failing means the order is never created. It fails closed: if risk is unreachable, nothing trades. There is no code path from a model output to an exchange that skips it.
- Rules
- 13 · all mandatory
- On failure
- Fails closed
- Bypass
- None exists
The order is placed the way the book allows
Placement is chosen from live conditions — passive when the spread is wide enough to earn, aggressive when it is not, split when depth is thin. Protective stops and targets go to the venue the moment the position exists, so a dropped connection is never an unprotected position. Realised slippage is measured against the decision price on every fill.
- Placement
- Adaptive per order
- Protection
- Resident at the venue
- Slippage
- Measured, not assumed
Every outcome is written down and kept
The trade closes and becomes memory: the conditions it was taken in, the reasoning at the time, the outcome, what went wrong and the lesson drawn. Rejections are journalled too — the record of what the system nearly did is usually more instructive than the record of what it did. Next time a similar setup appears, this is what gets consulted.
lesson: retest held — size normally
mistake: third retest in compression
lesson: retest held — size normally
recalled at the next similar setup
- Stored
- Context · reasoning · outcome
- Rejections
- Journalled as decisions
- Recall
- At the next similar setup
And the record tells you what is actually working
Results are decomposed by strategy, symbol, regime, session and hour, with cost drag broken out from gross performance. An equity curve tells you something is working; attribution tells you what — and usually that a comfortable overall profit is one symbol in one session carrying four that are not.
attribution, not a single curve
- Attribution
- 5 dimensions
- Costs
- Separated from gross
- Feeds back
- Into model calibration
Stage seven feeds stage three
Analytics is not the end of the pipeline, it is the input to the next one. Realised performance per regime is what recalibrates the arbiter's model weights, and the journal is what the analogue-recall model reads. The loop is the product — a system that trades without closing it is just a faster way to repeat yourself.
- Exchange→
- Analysis→
- AI→
- Risk→
- Execution→
- Journal→
- Analytics
- ↺back to 03