A complete futures trading system: it scans every liquid perpetual, measures what each one can actually deliver, solves the entry, target and stop that hit a chosen win rate, and executes — with a risk engine that halts it before you would have to. Built on 39 order books recorded continuously, not on downloaded candles. Trading MEXC USDT-M today; the venue layer is separate from the strategy, so more exchanges follow without rebuilding either.
Four stages, running continuously. Everything downstream is driven by one arithmetic identity, so the model can never be trained on different geometry than the bot trades.
A WebSocket recorder captures the full order book and trade tape across every selected perpetual, plus open interest and funding. This is data you cannot download after the fact.
Features are recomputed causally and tested with purged walk-forward folds, timestamp-clustered t-statistics and surrogate controls — so a result has to beat its own autocorrelation, not just white noise.
For each pair the engine solves the entry, take-profit and stop that reach a target win rate at a given leverage, after commission. Infeasible configurations are rejected rather than rounded.
The autopilot places the orders, manages the barriers, and enforces every limit: drawdown kill switch, daily loss cap, position count, trade cap, post-loss cooldown, stale quote and spread rejection.
Connect a MEXC key and watch it work. The dashboard is read-only by construction — the client it is built on has no order methods at all.
Pair selection, position sizing, entries, exits and halts are deterministic and tested. No human in the loop, and no language model in the trade path — the feed updates twice a second and anything that slow would be reasoning about a price that no longer exists.
A drawdown kill switch, daily loss limit, concurrent-position cap, daily trade cap and post-loss cooldown. Every plan is checked for liquidation distance and exchange minimums before an order is sent.
Equity, margin, open positions marked to market, and distance to liquidation — the number the exchange app buries. Updated every few seconds, from your own exchange key.
Encrypted with a key derived from your own password and held only while you are signed in. A stolen database or backup yields ciphertext. We never hold your funds and the key cannot withdraw.
The nightly report as it lands: information coefficients, family-wise significance bands, and the net figure after a 12bp round trip. You see the working, not a headline.
Real orders need a config flag, an explicit acknowledgement token and credentials, all three. Nothing turns it on automatically, and paper mode trades the same code against live prices without sending anything.
Out-of-fold, after commission, on a universe selected by traded value in the first month of the sample — because ranking by today's volume and testing backwards quietly deletes every coin that died.
Read straight from the bot's own output — no run recorded yet. Nothing on this page is typed in by hand, so it cannot drift from what the backtester actually produced.
Expectancy not yet positive after feesOut-of-fold only. In-sample curves are easy and meaningless; every trade shown was scored by a model that had not seen it, with purged walk-forward folds and a gap so no label leaks across the boundary.
Median price move by horizon on BTC_USDT, against the 12bp it costs to get in and out on MEXC's API fee schedule. Below the dashed line no amount of predictive power pays for the trade. Any system that will not show you this chart has not done the arithmetic.
Five strategy families have been tested and rejected. They are listed here rather than buried, because a system that shows only its wins is showing you a selection, not a result.
74 causal features, a from-scratch gradient-boosted tree, triple barrier labels, purged walk-forward CV. Out-of-fold AUC 0.50 — a coin flip. Rejected.
20 named hypotheses across 8 horizons and 40 symbols, market-neutral. Real effects, all smaller than commission. Best composite fell from t=3.55 in-sample to t=0.86 out. Rejected.
Significant at t=3.89 — until judged against phase-shifted fake moons instead of white noise, where p=0.084. Rejected, and the control changed for everything since.
Order-flow imbalance genuinely predicts the next ten seconds (IC +0.31, t +4.74). It is worthless anyway: BTC moves 1.4bp in a minute against a 12bp round trip. Short horizons are ruled out by arithmetic, not by a lack of signal.
The open question is whether microstructure state predicts moves hours ahead, where the move finally clears the fee. That is what the recorder is for, the instrument is built and validated against planted signals, and the answer publishes itself nightly. A result counts only when it clears a family-wise significance band and nets positive after fees — both, or it is not a trade.
Yes. It selects pairs, solves the geometry, places the orders and manages the exits without input. You can also run it in paper mode, which executes the same code against live prices and sends nothing.
Never. Your funds stay in your own exchange account. You connect an API key; we recommend one without withdrawal permission, and nothing here needs it.
On the dashboard, nothing but read. The client that serves your account panel physically has no order methods on it. Trading, if you enable it, is a separate path behind three independent gates.
Your stored exchange key is encrypted with it, so it cannot be recovered — you re-enter the key. That is the cost of the server not being able to read your secret, and we think it is the right trade.
MEXC USDT-M perpetual futures today. The exchange layer is separate from the strategy, so adding a venue does not touch the trading logic. One warning worth repeating: MEXC charges API-executed futures trades on a higher schedule than its app displays, and every figure on this page uses the API schedule.
No. Trading leveraged futures can lose your entire deposit quickly. No return is offered or implied, and past measurements do not predict future results.
Read-only to start. Everything measured, everything published — including the parts that did not work.