Running right now

Everything above describes a system that is, as you read this, watching the market. Below is its live output: for each coin, the model’s buy confidence this second, against the threshold it must clear to act. When a bar crosses its threshold, that is the bot deciding to enter.

The trajectory, coin by coin

The ticker is the snapshot; this is the path. Each dot is one prediction, plotted as the price the model saw against the confidence it produced — so you can watch a coin walk toward, or away from, its threshold over time.

The firehose from Act 1 is still flowing. The candles are still forming, the features still computing, the model still scoring. Everything on this page is one continuous loop — and the numbers above are its read on the market, this second.

The firehose

Start where the system starts: with what one coin actually looks like before anyone has interpreted it. The collector wakes up every second, asks Binance for the latest trades and the current order book, and writes a single row. Here are three of those rows for ETH.

The first ten columns are the familiar stuff: the symbol, the time bucket (ts_begin / ts_end as Unix seconds), open/low/high/close prices, a volume-weighted average price, the traded volume and the number of trades. Then it gets more interesting: the twb/tws columns are trade-weighted order-book depth — how much buy and sell liquidity sits within 0.3%, 1%, 3% and 10% of the mid-price. And those fourteen are just the beginning: another ~170 derived indicators are appended to every single row.

No human reads this. No model trains on it directly either — it is too fast and too raw. The first thing the system does is fold time into something a person, and a model, can actually hold: candles.

Folding time into candles

The first transformation is the oldest trick in market data: stop looking at every tick, and summarise each minute by four numbers — where it opened, how high it went, how low, and where it closed. A candle. Below are eight consecutive one-minute candles for ETH, aggregated from the same raw stream.

Anatomy of one candle

Each row of that table is one of these. The thin line — the wick — spans the highest and lowest price touched in the minute. The thick body spans open to close. A filled body means the minute closed higher than it opened; a hollow body means it closed lower.

The table, as a chart

Line those eight candles up in time and the table becomes something you can read at a glance: a quiet open, then a sharp sell-off on the minute with 14,016 trades, then a tentative recovery.

Candles are readable, but they are still just price. To predict, the system needs to turn this shape — and the order-book depth riding alongside it — into numbers that say something about what happens next. That is feature engineering.

Candles into signals

A candle tells you what just happened. A feature tries to say something about what happens next. On top of every k-line the system computes the full TA-Lib battery — roughly 180 indicators, including 61 candlestick-pattern detectors — and feeds them all to the models. Three are worth seeing up close.

Order-book depth & imbalance — the edge

What it is: how much real buy versus sell liquidity is resting within 0.3%, 1%, 3% and 10% of the mid-price — the twb/tws columns from Act 1.
Why it predicts: price history is the past; the order book is intent. A wall of bids just under the price tends to hold it up; a wall of asks above tends to cap it. This is the least textbook of the three, and the part that earns its keep.

RSI(14) — momentum

What it is: the Relative Strength Index, a 0–100 gauge of how one-sided recent moves have been. Above 70 is conventionally “overbought,” below 30 “oversold.”
Why it predicts: streaks exhaust themselves. Stretched momentum often precedes a pause or a snap-back, so RSI helps the model temper a signal that is running hot.

Bollinger Bands — volatility

What it is: a moving average with an envelope drawn a couple of standard deviations above and below it. The bands breathe with volatility.
Why it predicts: a tight squeeze means the market is coiled and a breakout is likely; price riding the upper band signals strength, hugging the lower band signals weakness.

So now there are ~189 numbers per second, per coin. What consumes them, turns them into a buy/sell confidence, and decides when to act? That is the machine.

The machine

Everything so far — the firehose, the candles, the features — is feedstock. This is the machine it feeds. The diagram below is the system’s own blueprint, recolored to match this page; it has two halves that meet at a single object, the model.

The Bintri optimization architecture. Box colour encodes node type (see legend): petrol = optimized, sand = artefact, clay = data or process. View full size ↗

The left half — building a model

A historical dataset is narrowed by filters (which coins, which times of day, how recent), producing the filtered dataset that model training learns from. Training also takes labels — derived from a forward time window (did the price move enough, soon enough, to be worth a trade?) — and a set of model parameters. The candles from Act 2 and the ~189 features from Act 3 are exactly what flows in here. Out comes a model, scored by a recall/precision curve and a backtest over held-out history.

The right half — trading it live

The grey Strategy box is the live loop. It holds a threshold per coin (how confident the model must be to act), an entry rule (what fraction of equity, after how many seconds’ delay), and logic to evaluate open positions for early closure. It consumes the latest k-lines in real time, and when the model’s confidence clears the threshold it sizes a position against the latest close / average fill price and closes the loop into Binance — futures and shorts included, trading fees accounted for.

One box on the left hides an enormous amount of work: model training. Choosing a good model means running it hundreds of times over different features, windows and parameters. How do you keep that honest and comparable?

How models are chosen

“Train a model” is really “train a few hundred and keep the best one.” Each run is a different combination of features, look-back and forward windows, label definitions and hyper-parameters. The only way to choose between them without fooling yourself is to log every run, identically, and compare. That is what MLflow is for.

The MLflow runs table for the trading models. Every row is one training run; every column a logged parameter or metric (precision, recall and the rest), so runs line up for direct comparison. View full size ↗

Each run records its parameters (what was tried), its metrics (how it scored — the precision/recall numbers behind that curve in Act 4), and its artifacts (the trained model itself). Because the logging is uniform, model selection becomes a sort-and-filter over a table rather than a matter of memory or vibes: you can ask “best precision at a given recall” and get an answer you can reproduce tomorrow.

The features from Act 3 are the raw material these runs consume; the winner is the model that gets promoted into the live Strategy box from Act 4. Which means it stops being a row in a table and starts doing something with real money.

Live orders on Binance

When the selected model’s confidence clears a coin’s threshold, the Strategy box does the thing all of this was for: it places an order. Not a paper order — a real one, on the exchange, with the account’s own funds. Here is a slice of that order history.

Executed orders on Binance. Each row is a real order the bot placed — market and limit, buy and sell, across XRP and DOGE — with its timestamp and fill price. View full size ↗

This screenshot is here to show one thing: the loop closes. Raw feed → candles → features → a trained, selected model → a strategy → an order that actually executes on a live exchange. It is not a performance claim. The system is experimental and has no proven profitability track to point at; what it has is a complete, working path from market noise to a real trade.

Which leaves one last question, and the most fun one to answer live: what is the model thinking right now?