Synapse Dynamix builds locally-run, neural-network trading systems for MetaTrader 5. The all-new in-house LiquidBrain maps the market's un-swept liquidity pools and predicts which one price sweeps first — a genuinely learnable direction read, validated by honest every-tick and walk-forward testing that refuses to ship anything it hasn't earned.
9.0 + the all-new LiquidBrain — live now · 0.5% risk · FTMO-safe
Pooled out-of-sample AUC for which liquidity pool sweeps first — far above the 0.50 coin-flip null. Then gated by MT5 every-tick and per-pair walk-forward.
These routes preserve the current production checkout, licensing, and support flows while the redesign keeps the technical brain showcase on the same page.
Every EA compiles from a single source — AgroAlgo9.mq5 — wrapped per product. Isolation is two-layer: physical (a wrong-product pair isn't in the signed catalog) and authorization (a per-product license endpoint gates the broker account).
The defining characteristic of AgroAlgo is not a secret indicator. It is an obsession with the gap between backtest and live, and an entire engineering apparatus built to measure it, close it, and refuse to ship anything that hasn't earned it.
The brain maps the market's un-swept liquidity pools — the nearest swing-high above and swing-low below — and predicts which one price sweeps first. That reframes an unlearnable up/down guess into a directional, structural read the model can actually learn (pooled OOS AUC 0.78).
Purged cross-validation with a null-shuffle test, then re-confirmed on MT5-native features, then an MT5 every-tick (1-minute OHLC) trading backtest, then per-pair walk-forward optimization — keeping only configs positive in both in-sample and forward, across two independent splits.
The 22 model features are computed inside MT5 itself by a shared exporter — the exact code that trains the model also runs live in the EA. There is no separate Python feature path to drift, so the live result tracks the backtest.
The brain is trained per instrument, exported to ONNX, cryptographically signed, and pushed over-the-air to client terminals. The EA runs inference locally on every bar. No central signal service — your VPS runs the model.
Each bundle is HMAC-SHA256 signed; the EA fetches its product catalog, verifies every file's hash, atomically swaps the active bundle, and rolls back / self-heals on any failure. Zero recompile to update the brain.
A single ONNX direction model per instrument, gated by a per-pair conviction threshold and closed with fixed stop and R:R exits. Every pair's entry settings are genetic walk-forward-optimized in MT5 — not one blanket setting bolted onto all of them.
Frequent retrains keep the brain regime-current — with the selection-bias risk itself guarded by the rigor harness. The signed ONNX bundles drive the MT5 EA directly — and because the brain is platform-neutral, new platforms can follow without retraining.
FTMO guards — news, weekend, session-flat — apply safety live, per client. One aggressively-validated brain serves a conservative prop client and an aggressive retail client. The client's EA settings decide the risk posture, not the model.
A complete technical walkthrough of the all-new LiquidBrain: how the liquidity-draw direction model works, how it evolved from 8.5, and how it was honestly validated. Every number traces to a real MT5 every-tick backtest or walk-forward window.
The iteration history is the strongest evidence of how AgroAlgo reaches quality. Each generation was earned — and the ones that failed on honest tests were thrown away, not shipped.
Full-history, embedded in the .ex5. The prior-generation multi-task, regime-routed neural network — the proven foundation 9.0 builds on.
Last fully-embedded build
Calibration heads, recency weighting. Honestly reverted to the 8.5 baseline after A/B — the discipline to discard months of work because the test said so.
Reverted · honestly
Signed OTA bundles, per-regime 9-ONNX ensembles, partial-TP runner, entry-mode dispatcher, M15 frequency brain, embargo true-OOS gates.
Proven production foundation
A lean, native rebuild — the new foundation, already a step up in profitability over prior versions. 15 instruments shipped and live over the signed OTA channel.
Live now
The breakthrough: raw up/down is unlearnable (~0.50 AUC), so the model instead predicts which un-swept liquidity pool price sweeps first — a learnable direction read (pooled OOS AUC 0.78). 22 features computed inside MT5, one ONNX model per pair, walk-forward-positive on every shipped pair.
Live now
Because honesty is the brand. AgroAlgo encodes its limitations into hard gates that are enforced by the system on itself — not footnotes hidden from the reader.
Raw up/down direction is genuinely unlearnable on this data (~0.50 AUC) — the label reframe to which liquidity pool sweeps first is what cracked it. What we ship is a real, modest-to-solid edge (silver and gold strongest, walk-forward OOS PF 1.14–1.77), not a hype number.
GBPUSD produced zero robust configs. Oil and Boom 500 were dropped too. A pair with no gate-passing model ships nothing rather than a loser — dark beats a losing model live.
Gold and silver are one metals bucket (~0.77 correlation) and the USD pairs share a dollar factor. The portfolio is diversified but not fully uncorrelated — size it at the portfolio level, not per pair.
Every number here traces to a real MT5 every-tick backtest and a two-split walk-forward. The brain is built, validated, and live now — and live forward evidence will still be published as it accrues.
Explore the LiquidBrain in full technical detail, or join the public Discord to follow the forward-test results as they accumulate.
