Everything it measures, it can also show.
Every screenshot below was taken on 15 September 2026 from the running application or from reports it generated. None are mock-ups.
Market viewer
Live order-book depth for Binance BTCUSDT futures: cyan is resting ask liquidity, magenta resting bids, the white line the midpoint. Every level of every update is drawn as it arrives, so walls appearing and pulling are visible second by second.
The panel at the bottom left is the part I used most: each venue's tick rate and how stale its last update is, so a slow feed can't masquerade as a price leader. Top right is the simulated trader's state.
Predictions in the viewer
The same viewer replaying a recorded hour with a trained predictor attached. At the bottom right is the 5-second predictor's current output with its running mean, spread and range. Under it, a grid of its 89 input features is coloured by how far each sits from its training statistics. The features are exchange-rate EMAs, RSI, ARIMA forecasts, cross-venue route deltas and volume skew.
The honest numbers: this model reaches a correlation of 0.17 and a relative absolute error of 99.6%, measured in-sample on data that includes this test hour. That is barely better than predicting no move, which is why predictions ended up as a filter around stronger triggers rather than a signal of their own.
Book service
The newest part: a pricing service for rolling BTC, ETH and SOL "up or down" markets over 5 and 15 minute windows, aligned to Polymarket's boundaries and settled against the same Chainlink feed Polymarket uses. It locks each window's price to beat, then prices both sides with a lognormal model on the live Chainlink stream, with order books from five exchanges wired in as source feeds.
Prices stream over WebSocket only when they move by at least one basis point, with a heartbeat otherwise. This board is a small page I wrote on top of that stream.
Maker playback workbench
Replays recorded books and draws a market-making strategy's own quotes over the real depth: magenta is recorded bid depth, cyan recorded ask depth, and the yellow and orange lanes are the lifetimes of the strategy's bid and ask quotes. The fill tape underneath records each fill with its slippage against the midpoint. It trains its fill, toxicity, direction and volatility models on the first four hours and tests on the fifth, so the strategy never sees its test data.
Eleven strategies and nine fill models are selectable, from "fill whenever touched" to a realistic queue model that only fills when the recorded book proves the queue ahead has gone.
Who leads BTC
730,934 top-of-book changes across eight feeds, scored for which venue moves first and which follows, with TCP round-trip probes from Cork to every endpoint so network position isn't mistaken for price leadership. Bybit spot came out as the driver; MEXC futures as the clearest follower.
Alpha search, cross-validated
A grid over lookback, horizon, simulated latency and trigger thresholds, trained on one period and validated on a later, separate one. Each row shows one-taker and round-trip markout in basis points, with a confidence figure, so the cost of paying to exit is always in view.
Maker simulation
A full account of one run: 17,513 ticks, 1,774 orders, 257 fills, inventory limits and markouts. Under the realistic queue model this expected-value-gated strategy ended an hour down 1.78 on a 10,000 account. It is shown here because it lost; that is what the simulator was built to catch.
Large-move audit
The thirty biggest BTC moves in a test hour, each traced across eight books. Bybit futures led most often; MEXC futures answered last, on average 411 ms behind. That gap is the opportunity the rest of the research tried to capture.