System Architecture

How everything connects

The core loop: Collect → Test → Learn → Improve

Binance
Live market data: candles, order books, funding rates
↓
Forward Engine
Active strategies · 37 coins traded · 5m→1D timeframes
Runs every 5 min, generates trades with order book proof
↓
SQLite Database
7,700+ trades · positions · regimes · decay metrics
Single source of truth: append-only trades, queryable
↓
Calibrator
Adjusts accuracy model
↓
Adaptive Bot
Picks winners, avoids losers
↓
User Backtests
Every backtest uses calibrated parameters from live data
More trades → better calibration → more accurate backtests

Multi-exit testing

Every strategy you run gets tested under multiple exit structures, not just one. The same signal, same data, different ways of closing the trade. The headline verdict uses the worst result so the honest read is always the first thing you see.

ATR 2:1
2x ATR stop, 4x ATR take-profit
Conservative default. Math puts breakeven win rate at 33%.
Fee-baked TP
1.5% net TP, 200-bar time stop, no hard SL
What our research lab uses. Favors mean-reversion signals.
ATR trailing
2x ATR initial, no fixed TP, trails 2x ATR below peak
Trend-follower structure. Lets winners run.
Your exit
Whatever you described in plain English
TP %, SL %, time, trailing, signal-based (e.g. exit when RSI > 70). All combinable.

Why multiple exits

Most of the question "does my strategy work" is actually "does my exit fit my signal." A mean-reversion entry needs small tight profit targets. A breakout entry needs a trailing stop that lets the trend run. Running all of them against your signal is the only honest way to tell you where your entry has edge and where it doesn't. If one exit works and another fails, that's your signal picking direction, not the strategy being broken.

Inside each exit: the regime heatmap

Ranging low vol
ADX < 25, quiet
Ranging high vol
ADX < 25, choppy
Trending low vol
ADX ≥ 25, smooth
Trending high vol
ADX ≥ 25, powerful

Every trade gets tagged with the market regime at entry (ADX for trend, ATR vs rolling median for volatility). Each exit model shows per-regime avg P&L so you can see where a strategy works and where it doesn't. Example: a mean-reversion entry might make +0.3%/trade in ranging low-vol under fee-baked exits but lose in trending high-vol. That's an actionable prescription: add a regime filter or avoid trend days.

Verdict tiers

CONFIRMED
Worst-case exit is still profitable on unseen data. Strategy holds up across exit structures.
MIXED
Worst exit loses, but some exit structure makes the signal profitable. Entry has edge; the exit choice matters.
DEBUNKED (signal evidence)
No exit is profitable, but one shows win rate ≥ 50% near breakeven. The signal picks direction; costs + exit structure are the gap to profitability.
DEBUNKED
Win rate near the 28-30% structural floor (our ATR 2:1 math breakeven is 33%). Signal doesn't pick direction. Usually needs a different entry, not a different exit.

Every result shows the walk-forward split (60% train / 40% test) per exit model so you can check that in-sample performance survives out-of-sample. Overfit strategies get flagged automatically.

Owner only