Live · running in production

MIDAS

Market Intelligence & Directional Analysis System

Not an indicator. Not a strategy. An intelligence core. The way a language model reads language, MIDAS reads the market — candle by candle. It learns from a decade of history, keeps learning as the live market moves, and answers one question with a number: which way, and how sure.

An SS7 Trader research program
Resolution
1 min
Probability
0.71
Measurements / candle
23
0
Years in the market
0
Years of history
0
Bars trained on
0
Charts covered
0
Research programs
The core

An intelligence, not a product.

Most trading tools are a set of rules someone wrote down. MIDAS is a trained intelligence. Nobody told it that a particular candle shape means anything — it was shown a decade of market history and learned which conditions actually preceded a move, and which only looked like they did. And it does not stand still: as the live market adds new history, the core is re-trained and re-validated, so it moves with the market instead of ageing against it. It also reads each moment across several time horizons at once, so a trader is not locked to one. Every SS7 Trader system runs on this same core.

It learns, it is not configured

No thresholds tuned by hand until a backtest looks good. The model is fitted on history, re-validated as the market moves on, and only then released.

It knows what it does not know

Silence is a valid output. Below its floor, MIDAS publishes nothing rather than guessing.

Perception

Twenty-three measurements on every candle.

A price on its own means very little. MIDAS reads each bar through four independent lenses at once, and it is the agreement between them — not any single one — that produces a signal.

Market structure
Where price has turned before, and whether those levels still hold.
Momentum & volatility
How hard it is moving, and whether that size is normal for this hour.
Participation
How much activity actually sits behind the move.
Session
The clock, in UTC, so Tokyo and New York are not one market.
The individual measurements, their weights and the thresholds that act on them are not published.
Provenance

A decade of market history, audited first.

A model is only as honest as the data underneath it. Before any training begins the history is normalised to UTC and checked for the duplicates, broken bars and gaps that quietly corrupt a result.

0
1-minute bars loaded
0
Duplicate timestamps
0
Invalid OHLC rows
0
Adjacent-minute continuity
0
Charts audited
EET → UTC
Timezone normalised
Read from the training run's own data-quality report, not typed in by hand.
The MIDAS standard

Tested until it stops, not tuned until it looks good.

Any model can be made to look brilliant on data it has already seen. That is the easiest thing in this field to do, and it is why so much of what is sold to traders falls apart in the first live week. Five checks, in order, fixed before a model is ever trained.

01
Step 01

A sealed holdout, opened once.

A block of history is locked away before training begins. It never chooses a model, a threshold or a setting — it is opened at the end, and it answers one question: does this work on time it has never seen? If a model only performs on data it helped select, this catches it, and nothing else does.

TrainThe model learns here — 70% of the history.
ValidationSettings and thresholds are chosen here, and only here.
SealedLocked before training. Opened once, at the very end.
Boundaries are chronological, never random. Shuffling minutes out of the middle of a training set lets a model see its own future.
02
Step 02

Walk-forward, not one lucky window.

Instead of a single split, the model is retrained and re-tested repeatedly forward through time, the way it would actually meet the market. A result that survives every window is worth something. One that appears in a single period is not.

Trained on — everything before the cut Judged on — only what came after it
Each pass moves the cut forward and repeats. A model that survives every window is worth something; one that appears in a single window is not.
03
Step 03

Nine gates, written down first.

The conditions a model must meet are fixed before the test, not chosen afterwards to fit whatever came out. A model that misses one does not ship, however good the headline looks — and that decision has been made more than once.

Gates are set before a single result is seen. Moving one afterwards is how a failed model becomes a marketed one.
04
Step 04

Microstructure integrity.

Markets contain conditions that look like enormous predictive edges and are really artefacts of how a broker quotes price. Give a model an unguarded window like that and it will report a number no honest system can reach. MIDAS screens for them explicitly: an edge that exists only inside a widened spread is removed before it can reach a trader.

Quoted spread — what a trade has to cross Size of the move — what the model is predicting
Schematic. Where the spread towers over the move, any “edge” found there lives inside the quote rather than in the prediction — so those hours are excluded before training.
05
Step 05

Verified before it ships.

The model that runs on your machine must be arithmetically identical to the one that was validated. Every deployed core is checked against its research version, decision for decision, before release.

model graph check pass
research vs deployed — max deviation 5.96e-08
client compile 0 errors, 0 warnings
regression tests 4 passed, 0 failed
covering causal features · exact expiry across gaps · UTC session flags · timezone conversion
Built by SS7 Trader

Eleven years in this market.

SS7 Trader has worked in this market for eleven years, and has built a name on one thing: tools that behave the same way in a live account as they did the day they were demonstrated. MIDAS is that experience turned into research — no licensed signal, no white-labelled engine, no third party deciding what our traders see.

Eleven years

Continuously active since 2015.

Own research

Every model trained in-house, on data we collect ourselves.

Own infrastructure

Our feeds, our servers, our pipeline, end to end.

A real team

Researchers and engineers, reachable and accountable.

Coverage

Where MIDAS runs today.

The core is market-agnostic, but it is only released on a market once that market has passed the same protocol. Forex is live now. Gold and the rest are in active development and will be announced when — and only when — they qualify.

Forex
Live
Gold
In development
Silver
In development
Crypto
In development
Indices
In development
Stocks
In development

The hardest part was never building a model.
It was learning which ones to throw away.

Thunder AI

The first system built on MIDAS.

MIDAS is the core. Thunder AI is what it looks like in a trader's hands — live signals across the Forex market, delivered where you trade.