Strategies

The strategy stack

name	universe	timeframe	entry	exit	status
`mean_reversion_v1`	BTC/ETH/SOL/AVAX/DOGE	15-min	EMA-9 crosses below EMA-21 + RSI(14) < 35	RSI > 65 OR hold > 4h OR stop hit	🚫 negative edge on all windows — 270d: 100 trades, 0.73:1 W:L, +$1.45
`mean_reversion_v2`	BTC/ETH only	15-min	EMA-9/21 cross-down + RSI<32 + not in downtrend	12h max OR stop>4% OR profit-lock>+1.8%	🚫 270d: 20 trades, 0.86:1 W:L, +$0.22 — wider stops help a bit, not enough
`momentum_v1`	BTC/ETH/SOL	15-min	fast_ema > slow_ema * 1.0005 + close > 16-bar high	2.5 ATR stop, 24h max, trend-break exit	🚫 270d: 66 trades, 0.72:1 W:L, -$3.41 — win rate 58% but stops kill winners
`regime_mom_v1`	BTC/ETH/SOL	15-min	momentum_v1 entry + regime must be `trending_up` (ADX ≥ 20 + EMA slope > 0)	2.5 ATR stop, 24h max, trend-break exit	✅ 270d: 29 trades, 1.88:1 W:L, +$60.30, sharpe +0.535
`regime_mom_v2`	BTC/ETH/SOL	15-min	regime_mom_v1 entry + `volume > 1.2 × 20-bar SMA` (the load-bearing fix)	2.5 ATR stop, 24h max, trend-break exit	✅✅ 365d: 13 trades, 1.85:1 W:L, +$33.99, sharpe +0.898 — FIRST to pass strict 0.5 Sharpe gate on 365d
`regime_mr_v2`	BTC/ETH	15-min	mr_v2 entry + regime must be `ranging` (ADX ≤ 12)	same as mr_v2	🚫 90d: 2 trades, +$0.16 — too few trades for stat-sig

Why regime matters: classical retail crypto strategies fail in isolation because they’re applied to wrong regimes. Mean-reversion bleeds in trending markets (stops fire before bounces play out); momentum bleeds in ranging markets (trend-break exits fire on noise). The regime filter cuts out the regime-inappropriate trades and stabilizes the backtest across multiple windows. regime_mom_v1 90d → 270d → 365d all positive (54 trades, 1.19:1 W:L on 365d) is the load-bearing structural-edge number.

Why two strategies? Single-strategy bots die when regimes rotate. BTC trades in clear regimes (trending / ranging / volatile) that change every 3-7 weeks. The bot needs to know which regime it’s in and call the appropriate strategy.

mean_reversion_v1 — dead on arrival

The v1 strategy shipped on day 1 and immediately failed its 30-day backtest:

9 trades over 30 days on real BTC/ETH/SOL/AVAX/DOGE bar data
55.6% win rate (sounds OK; isn’t — see below)
$-1.94 net P&L
0.27:1 win:loss ratio (every win made $0.20, every loss cost $0.73)
sharpe-lite: -0.24

The problem is structural: the 2.5% hard stop is tighter than the average upside capture in a trending regime, so the bot cuts winners short and lets losers run. This is the textbook mean-reversion death-by-tight-stops. The strategy doesn’t have edge — its wins are smaller than its losses.

Live-fire status: blocked. The harness tests/test_strategy_backtest.py is the gatekeeper; v1 never produced a positive verdict and so v1 doesn’t get a Trader.place() call into the live paper account.

momentum_v1 — spec only

Still drafting. The intent: breakout in the direction of the trend with a tight trailing stop. Trade WITH momentum, not against it.

entry:  price makes a new N-bar high AND ADX > 25 (trending) AND above 200-EMA
stop:   1.5 ATR trailing
exit:   stop hit OR new N-bar low (trailing stop trails below)
universe: BTC, ETH (most liquid pairs only)

This is the strategy v1 was not. Backtest pending.

regime_classifier — the controller

A single strategy is not enough. The regime classifier decides which strategy to call per pair per bar:

trending_up   → momentum_v1
ranging       → mean_reversion_v1 (with relaxed stops, since stops are the v1 killer)
trending_down → momentum_v1 (shorting, but Alpaca retail crypto is long-only — handled below)
high_vol      → pause new entries, sit in cash

Alpaca retail crypto constraint: long-only. So in trending_down the bot can either (a) sit in cash (raising cash%), or (b) short a stablecoin pair, or (c) buy inverse contracts. Strategy must respect venue constraints.