Skip to content
777 Raptor

Blog · EMIL

How EMIL learns without being retuned

EMIL adjusts its own inputs from what the market is actually doing, so nobody has to re-optimise it every quarter. Here is the mechanism, and where it stops.

18 June 2026777 Raptor2 min reademilmethodadaptation

Most automated trading logic is a snapshot. Someone chooses the inputs, tunes the weights against a period of history, ships it, and the market moves on. The logic keeps applying yesterday's relationships with yesterday's confidence, and the only thing that changes is the equity curve.

EMIL is built the other way round. It adjusts its own inputs from what the market is currently doing, continuously, without a human re-optimising it.

What "learns by itself" actually means here

It means four specific mechanisms, not a vague claim about intelligence.

It scores its own inputs against outcomes. EMIL watches a set of features — price, spread, realised volatility, session state, exposure, cross-asset relationships — and tracks how well each one has been describing what subsequently happened. A feature that has been informative gets weighted up. One that has stopped being informative gets weighted down. Nobody edits a config file to make that happen.

It notices when a relationship breaks. A correlation that held for six months and inverted last week is not averaged into a comfortable middle. EMIL tracks stability across sub-windows and treats a broken relationship as broken, which is the opposite of what a static model does.

It classifies the regime before it acts. Trending, ranging, expanding, contracting. The same conditions that make one reading useful make another useless, so the classification comes first and gates everything after it.

It re-ranks continuously, not on a schedule. There is no quarterly retraining window during which the system is knowingly stale.

What it does not mean

It does not mean EMIL learns its way to a profit. Adaptation reduces the chance of acting on a dead relationship. It does not remove market risk, and a system that adapts faster is not a system that is right more often.

It also does not mean the model is a black box you have to trust. Every reading EMIL produces is logged with the inputs that produced it, in language a risk committee can read:

Example
09:38:44  OBSERVE   Regime shift detected: EURUSD trending → ranging.
09:40:12  ADAPT     Cross-asset correlation weight 0.31 → 0.11 (relationship unstable 9d).
09:40:18  PROTECT   Exposure 62% of mandate.
09:41:02  REFUSE    Order rejected: would breach concentration limit.

That fourth line matters more than the first three. A system that explains its refusals is a system you can audit.

Why adaptation and authority are separate

Learning changes what EMIL thinks. It never changes what EMIL is allowed to do. Those are deliberately different subsystems:

  • Adaptation happens inside the intelligence layer.
  • Authority is a mandate you wrote, confirmed by typing a confirmation word.
  • Enforcement happens outside the intelligence layer entirely, in the risk engine, in the order path.

So EMIL can conclude that a larger position is warranted and still be refused by the exposure limit — and the refusal is logged. Nothing EMIL learns can widen its own boundaries. That separation is the whole design.


Next: what EMIL actually does once you arm it.

Trading leveraged products carries a high level of risk to capital. Technology can improve analysis and controls; it cannot remove market risk.