One brain that thinks the order flow

Earlier versions ran separate logic per panel: the Dashboard computed one thing, the Heatmap another, the NT8 strategy a third. The result was conflicting readouts — green in one window, red in the next. ONE CORTEX is exactly one brain: a self-learning Long Short-Term Memory (LSTM) network that learns from the order book and the tape of the E-mini S&P 500 and hands the identical read to every surface.

This article explains how the one brain is built, what its heads predict, what it learns from — and how it honestly shows whether it is any good right now.

One brain, not many separate signals

There is one central model. It makes the call, and the Dashboard, Heatmap, DOM, Journal and strategy read exactly the same output. No panel contradicts another, because there are no parallel second opinions.

One brain with several heads is more consistent than many separate detectors overruling each other. Every head sees the same features — each one answers a different question.

The heads on one backbone

The heads sit on the same LSTM backbone. Each is trained on its own clearly defined target — none of them is an opinion, every one is answered by the market after the fact:

One more head measures how often a signal flips before a trade resolves. It deliberately runs as a measurement only and influences no decision. The Swing readout on the Dashboard is not a learned head of its own: it combines regime and trend direction and shows the target zone as a VWAP band with an invalidation price.

Features, backbone and what CORTEX learns from

At every time step ONE CORTEX processes a vector of several dozen microstructure features: delta over several windows with acceleration and divergence, order-book imbalance across several levels, walls, pulls and refills, absorption and sweeps, iceberg and spoof traces, large prints, tape speed, distance to VWAP, value area and VPOC, Wyckoff phase and gamma-pin traces. Deliberately no clock time — the model should read the market, not the calendar.

A sequence covering roughly the last minute and a half runs through two stacked LSTM layers and a compact dense block, with the heads on top. Training runs in its own background process, so the live signal never waits for learning.

The same read everywhere — Dashboard, Heatmap, DOM, strategy

The app produces exactly one signal and distributes it to every window. That is the real point of ONE CORTEX:

• Dashboard. SCALP and SWING readouts with confidence, plus Bid/Ask Pulse — the real-time pressure gauge of aggressive buyers and sellers.
• Heatmap. This is where CORTEX runs and learns. The signals are embedded visually, together with sweep/absorption, reversal and event pills.
• CORTEX DOM. The PRESS HUD at the inside market shows who is pushing and who is winning.
• NT8 strategy (Pro). The Cortex OEM AI Strategy receives the same read about ten times per second. The direction comes from CORTEX alone; an entry additionally needs enough confidence, confirmation from the order book and the gate of the target-odds head. If CORTEX flips firmly to the other side before the first target, the strategy goes flat.

Signal format (simplified)

{
  "v": "cortex-signal/0.2",
  "seq": 48213,
  "modelAge": { "sessions": 6, "samples": 41870 },
  "regime": { "label": "TREND", "conf": 0.61 },
  "scalp": { "dir": "LONG", "conf": 0.58, "ttlSec": 30, "entryReady": true, "exitCue": "NONE" },
  "swing": { "bias": "BULL", "targetZone": [5412.25, 5418.50], "invalidation": 5403.75 },
  "bracket": { "pLong": 0.54, "pShort": 0.41, "live": true },
  "provenance": { "health": { "balAcc": 0.37, "mcc": 0.06 } }
}

New in 0.5.30: CORTEX DOM and CORTEX LEARNING

Two additions make the CORTEX read easier to follow in real time. CORTEX DOM puts the pressure around the inside market into one compact display, and CORTEX LEARNING shows — step by step — that the model is learning from your own feed.

None of these displays is a trading signal on its own and none of them promises an edge. CORTEX DOM makes the order flow at the inside readable; CORTEX LEARNING makes the learning visible. What CORTEX learns from, and how fast, depends on your own Level-2 feed.

Honest, not green — self-test and learning curve

A learning system does not always have an edge — and ONE CORTEX does not pretend it does. Two readouts say openly where it stands. Both measure the quality of the model, not the market direction — they are not a buy or sell signal.

Model self-test

Based on the balanced accuracy and the Matthews correlation coefficient (MCC) of the predictions:

Learning curve per trading day

The learning curve checks whether CORTEX gets better over the days instead of just counting sessions: for each trading day, the hit rate of the target-odds head is measured against the trivial baseline — using predictions that were fixed before the market answered them. If the 95 % band lies entirely above zero, the verdict is EDGE; entirely below, BELOW; in between, OPEN.

Both readouts stay visible, even when they are uncomfortable. A tool that honestly reports WEAK, OFF or BELOW is worth more than one that always glows green.

What ONE CORTEX is NOT

• Not a ChatGPT wrapper. No large language model — a dedicated LSTM with numerical inputs and numerical outputs.
• Not a black-box oracle. Every signal, every output and the Cortex Ledger are traceable.
• Not a guaranteed edge. ONE CORTEX promises no performance and no profits. When no edge is measurable, the self-test and the learning curve show exactly that.
• Not an autonomous bot without controls. Auto-trading exists only with the NT8 strategy in Pro and Ultra — with a daily limit, pause and flat-all right in the chart.

Foundations and testability

ONE CORTEX is built from well-studied components, not from a secret recipe. Whether they hold up in the market is decided by data: pre-registered and public, even when the result goes against us.

The components and their sources

• A sequence core of LSTM layers. It reads the order book as a sequence in time, not as isolated snapshots. Hochreiter & Schmidhuber (1997), Long Short-Term Memory, Neural Computation 9(8).
• Learning with a replay buffer. Each head trains from a buffer of stored examples so that consecutive, very similar market moments do not skew the learning. Mnih et al. (2015), Human-level control through deep reinforcement learning, Nature 518.
• Order-book features, not just price. Pressure and imbalance in the book carry more information than price alone. Cont, Stoikov & Talreja (2010), Operations Research 58(3); Lehalle & Laruelle (2018), Market Microstructure in Practice; Sirignano & Cont (2019), Quantitative Finance 19(9).
• Your model, your feed. CORTEX starts without a pre-trained model and learns on your machine from your own feed. It learns trade outcomes only from the strategy's trades, never from manual, test or copied trades.

How we test

• Pre-registered. Question, data set, metric and threshold are fixed before anything is computed.
• Against chance. The baseline is always chance, in the model self-test above as well as in every pre-registration.
• Public. Results, negative ones included, are in the lab notebook.

CORTEX is an analysis and decision-support tool, not investment advice. Trading futures carries substantial risk of loss, up to the total loss of the capital employed.

Frequently asked questions

Related articles