Strategies that show
their working.
A quantitative strategy pack for Gunbot, measured on three years of real hourly data across seven markets after fees and slippage — and an AI desk that can propose orders into it without ever touching your exchange keys. Every defect the audit finds is published, including the ones that make a strategy look worse.
Two things, and the evidence for both.
The strategies
Market making, mean reversion, trend following, statistical arbitrage, grid accumulation and scalping — built on published models and calibrated for crypto. They run inside a Gunbot you already own.
- 34 documented, every configurable setting explained
- Measured on real market data, reported at the median across markets
- One log format, one colour contract, one side-panel grid
The AI desk
Connect any model to your own Gunbot. The desk screens your enabled pairs, asks your model to pick one, runs the answer through a fail-closed risk gate, and delivers the surviving order as a signed webhook, a REST call, or both.
- Propose, never dispose — the model picks from a filtered list, by index
- Exchange keys never enter the desk; they stay in the bot you control
- Fail closed everywhere — a missing limit blocks the order, never guesses
Your keys never leave your bot.
Both products act on a Gunbot you run. Neither holds exchange credentials, and neither needs a cloud service to be up for your bot to keep trading.
The unflattering numbers are published too.
Every audited strategy is run across seven markets and three years of real hourly data, after a 0.10% taker fee and 2 bps of slippage, and reported at the median across markets — never the best pair. The buy-and-hold column sits next to the return column, because most of these strategies lose to simply holding the asset over this window, and the report says so.
| Strategy | Median return | Buy & hold | Edge | Max DD | Invariants | Fixes |
|---|---|---|---|---|---|---|
| kalmanGrid | 0.3% | 39.2% | -38.5% | 0.3% | 5 / 5 pass | 4 |
| meanRevOscillator | 6.1% | 39.2% | -42.8% | 2.2% | 5 / 5 pass | 3 |
| rangeRider | 4.3% | 39.2% | -20.1% | 2.9% | 5 / 5 pass | 2 |
| volHarvester | 3.4% | 39.2% | -34.6% | 1.4% | 5 / 5 pass | 4 |
| volRegimeMM | 3.4% | 39.2% | -34.6% | 1.4% | 5 / 5 pass | 4 |
| WaveTrend_LB | 1.1% | 39.2% | -46.4% | 12.2% | 5 / 5 pass | 4 |
58 dated entries, back to the first sweep.
Every fix that landed, written in plain language when it landed — 47 patch notes, 5 pack-wide milestones, and one entry per strategy as its audit completes.
What none of this claims.
An audited strategy is correct, not profitable. Nothing here was measured on a live exchange. The measurement engine models spot and long-only with a single price per bar, so it carries no order book and cannot measure a strategy whose edge is the spread. Three years is one history, and a strategy tuned on it can be fitted to it.
No strategy in this pack should be run unattended on capital that matters. The reports state the window, the market count and the sample size for exactly that reason.