Regime-Matched Multi-Strategy Portfolio Architecture | Generated by AI

Home 2026.06

Question: Explain the architecture and logic of my multi-strategy trading bot — three strategies across five instruments with a shared risk layer.

Answer:

Your bot is a classic regime-matched portfolio: each strategy is paired to the statistical behavior of its market, and a shared risk layer normalizes them all into comparable bets. Here’s the system layer by layer.

1. Why each strategy fits its instrument

The core insight is autocorrelation of returns at different timescales:

You’re essentially diversifying across signal types, not just tickers. Mean reversion and trend following have historically low or negative correlation of returns — when chop kills the trend strategies, the reversion strategies eat; when a strong trend kills reversion, the trend strategies eat.

2. The three signal engines

Mean reversion (z-score):

def zscore_signal(closes, n=20):
    mu = closes[-n:].mean()
    sigma = closes[-n:].std()
    z = (closes[-1] - mu) / sigma
    if z <= -2: return "BUY"    # 2σ below mean: statistically stretched
    if z >= 0:  return "SELL"   # reverted to mean: edge is spent

The asymmetry matters: you enter at −2σ but exit at 0, not +2σ. You’re harvesting only the reversion, not betting on continuation past the mean. That keeps holding periods short and avoids overstaying.

Momentum breakout (Donchian):

def donchian_signal(highs, lows, close):
    if close > highs[-21:-1].max(): return "BUY"   # breaks prior 20-bar high
    if close < lows[-11:-1].min():  return "SELL"  # breaks prior 10-bar low

The asymmetric channel (20 in / 10 out) is deliberate: slow to enter (confirm the breakout is real), faster to exit (give back less when the trend dies). This is the classic Turtle-style structure.

Trend following (dual SMA):

def sma_trend_signal(closes):
    fast, slow = closes[-20:].mean(), closes[-50:].mean()
    if fast > slow and closes[-1] > slow: return "BUY"
    if fast < slow:                        return "SELL"

The extra condition price > slow SMA filters out entries during a crossover that happens while price is already rolling over — a cheap whipsaw reducer.

3. The risk layer — where the real engineering is

ATR-based 1% sizing is volatility normalization:

risk_dollars = 0.01 * equity
stop_dist    = 2 * atr14
shares       = risk_dollars / stop_dist

Since stop_dist scales with volatility, a quiet instrument (GLD) gets a large position and a violent one (IBIT) gets a small one — but both lose exactly 1% of equity if stopped. Every trade is the same-sized bet in risk space. This is what makes a 5-instrument, 3-strategy book coherent: positions are comparable units.

GTC stop at the broker decouples risk enforcement from bot uptime. Since you run on GitHub Actions (which can be delayed, rate-limited, or down), the stop living server-side at the broker means the 1% cap doesn’t depend on your cron firing. This is the single most important reliability decision in the design.

Correlation filter (max 2 longs in SPLG/QQQM/IBIT): those three are all risk-on assets — in a risk-off shock they sell off together, so three “independent” 1% bets are really one ~3% bet. Capping at 2 bounds the correlated drawdown.

No leverage, 95% cash cap, 25% per-instrument cap: hard guardrails so that no sizing bug, ATR collapse (tiny ATR → huge share count), or signal pile-up can blow past sane exposure. The 25% cap specifically protects against the ATR-sizing failure mode where low volatility produces an enormous position.

4. Execution model and its edges

The engine wakes every 15 min, evaluates all five, and acts only on fresh signal transitions — i.e., it’s edge-triggered, not level-triggered. That prevents re-entering the same signal every cycle.

Things worth keeping an eye on in this design:

  1. Bar-close alignment: your 1h (IBIT) and 4h (GLD/USO) strategies only have new information once per 4–16 cycles. Make sure the bot evaluates them on closed bars only, or the SMA/Donchian values will flicker intra-bar and generate phantom signals.
  2. Gap risk: a GTC stop caps risk at the stop price, not at 1%. An overnight gap through the stop (especially IBIT tracking 24/7 BTC against US-session trading) can lose more than 1%. The 25% position cap is your real worst-case bound here.
  3. GitHub Actions jitter: scheduled workflows can fire minutes late or get skipped under load. Since stops are at the broker, this only delays entries/exits on signal, not risk enforcement — which is the right failure mode to have.
  4. Z-score with σ→0: in dead-quiet sessions the 20-bar std can get tiny, making z-scores explode. A minimum-σ floor avoids spurious −2σ entries on noise.

Summary in one line

Three uncorrelated edge types (fade / breakout / trend), each matched to its market’s autocorrelation regime, all flattened into identical 1%-risk units by ATR sizing, with broker-side stops so risk holds even when the bot doesn’t run.

One caveat since this is live money: I’m not a financial advisor and nothing here is a claim that these edges are currently profitable — the explanation covers the mechanics and engineering soundness, which are solid. The empirical question (do these signals still carry edge net of fees/slippage at 15-min granularity?) is only answerable from your backtest and live fills.

Back Donate