Platform Overview

Regime Trader is a multi-strategy automated trading systemthat trades equities via an Alpaca brokerage account. Three independent strategy “sleeves” share the same account with configurable capital splits (default 60% HMM / 40% Brando / 10% Short Hedge):

An independent risk managerenforces hard limits across all sleeves: circuit breakers, drawdown caps, and volatility caps. A symbol cannot exist in both the HMM and Brando pools simultaneously — the system enforces disjoint universes.

App Pages

Trading Schedule

All times are Eastern (ET). The scheduler runs independent jobs on weekdays only, plus quarterly maintenance tasks.

HMM Volatility Strategy

The HMM engine uses a Gaussian Hidden Markov Modelwith full-covariance matrices to classify market volatility environments. It operates strictly as a volatility classifier — it does not predict price direction. The regime state informs an allocation layer that adjusts position sizes, risk limits, and leverage.

Model Selection

• Candidates tested: n_components ∈ [3, 4, 5, 6, 7]
• Each candidate trained with 10 random restarts (keeps highest log-likelihood)
• Winner selected by lowest BIC (Bayesian Information Criterion)
• States ranked by expected volatility and assigned human-readable labels

Regime Labels & Allocation Floors

• BULL / EUPHORIALowest volatility states. Prices trend positively or drift safely. Full allocation (95%) with up to 1.25x leverage.
• NEUTRAL / BEARIntermediate volatility. Market is ranging or in a moderate downtrend. Cautious exposure (max 30%), leverage disabled (1.0x).
• DEEP BEAR / CRASHHigh volatility with rapid moves and gap-downs. Allocation forced to 0% (full cash). These regimes override all other signals.

Filtered Inference (No Look-Ahead)

The system uses the forward algorithm exclusively for live inference. Probabilities at time t depend only on observations up to and including t— no Viterbi, no smoothing. This prevents any look-ahead bias and is enforced by automated tests.

Uncertainty & Flicker Detection

• Uncertainty mode: if max regime probability < 55% or flickering is detected, allocation is halved and leverage forced to 1.0x
• Flicker rate: counts regime changes per 20-bar window. If changes > 4, the model is “flickering” and risk is reduced
• Confirmation: a regime must persist for 3+ consecutive bars before it's considered confirmed

Order Execution

HMM orders use whole shares only (fractional quantities are floored). Entry is submitted as a DAY limit at a slightly aggressive price; if unfilled within 30 seconds, the remaining quantity is submitted as an OPG market order(fills at next morning's opening auction). After fill, a GTC stop-loss is attached as a standalone order so it persists overnight until cancelled by the next rebalance cycle.

Regime Allocation Strategies

Each regime is assigned a strategy based on its volatility rank. The strategy controls allocation sizing, leverage limits, and stop/target distances.

Brando Level Trader

A swing-trading strategy that trades off confirmed support and resistance levels. The core philosophy: identify where institutional interest clusters (levels with multiple touches on high volume), wait for a rejection signal, enter with predefined risk, and manage the trade through staged exits.

Level Detection

• Scans the last 120 daily bars (~6 months) for swing pivots
• Nearby pivots within 0.5x ATR are clustered into a single level
• A level is confirmed when it has ≥2 touches on above-60th-percentile volume
• Levels are scored 0–1 based on touch count (30%), duration (25%), recency (25%), and volume (20%)

Entry Criteria (all must hold)

1. Price is within 0.75x ATR above a confirmed support level
2. All enabled macro gates pass (see Advanced Features below)
3. Last completed daily bar is a rejection candle:
   • Low tested the level (within 0.2x ATR)
   • Close recovered above the level
   • Body ratio ≥ 0.3% (not a doji)
   • Volume ≥ median of lookback window
4. Risk sizing produces a valid quantity ≥ 1 share

Position Sizing (Size-to-Zero)

Exit Rules (Priority Order)

Target Calculation

• TP2: nearest confirmed resistance above entry, or entry + 3.5 × (entry − stop) as fallback
• TP1: entry + 50% × (TP2 − entry) — the halfway point

Brando Advanced Features

All of the following are independently toggleable via sleeve params. They default to conservative (off or permissive) and can be enabled incrementally.

Macro Gate (B2)

Three market-regime filters that block new Brando entries when macro conditions are unfavorable. All gates default to permissive when data is unavailable to prevent infrastructure outages from halting trading.

Multi-Timeframe Confirmation (MTF)

Disabled by default (mtf_confirmation_enabled=0). When enabled, a daily support signal is only valid if there is also a confirmed weekly support levelwithin 1.5× the weekly ATR. Expected effect: ~40% fewer signals, ~10 percentage-point win-rate improvement.

Chandelier Trailing Stop

Disabled by default (chandelier_enabled=0, multiplier chandelier_atr_multiplier=2.0). After TP1 is hit, instead of a static breakeven stop, the stop trails as:

This lets winners run in trending markets while still locking in gains from TP1.

Correlation Guard

Disabled by default (corr_guard_enabled=0, window corr_guard_window=60). When enabled, skips a new entry if the candidate symbol has a 60-day return correlation ≥ 0.70 with any already-open Brando position. Prevents over-concentration in correlated names.

Brando Configuration

All parameters are stored in sleeve_config.params for the BRANDO sleeve and can be changed via the Bot Engine page or API.

News Screener

The screener runs daily before market open (08:30 ET) and can be triggered manually from the Bot Engine page. It identifies symbols “in play” by analyzing recent financial news. News tells you where to look. Levels tell you when to trade.The screener feeds the Brando watchlist — it does not generate trading signals.

News Sources

• Alpaca News API — included with your trading credentials, returns articles with tagged ticker symbols
• Finnhub — free API (requires FINNHUB_API_KEY in .env), extracts tickers from headlines

Screening Pipeline

1. Pull articles from all configured news sources (last 24h, up to 200 per source)
2. Count symbol mentions across all articles, excluding symbols already in any pool
3. Take top 30 candidates by raw mention count
4. Filter: must be tradable on Alpaca, avg daily volume > 500K, price $10–$500
5. Score remaining candidates (composite 0–1)
6. Return top 10 as pending suggestions

Scoring Formula

Screener Parameters

Risk Management

The risk manager operates independently of both strategies. It enforces hard limits that cannot be overridden by signals. On any HMM or data error, the system defaults to risk-off (hold current positions, no new trades).

Circuit Breakers

Position Limits

• Max portfolio exposure: 100% of sleeve equity
• Max single position: 95% of sleeve equity
• Max leverage: 1.25x (only in low-vol bull regimes)
• Max risk per trade: 1% of portfolio

Global Volatility Cap

If 20-day realized volatility exceeds the 95th historical percentile, max allocation is capped and leverage is disabled regardless of HMM output.

Portfolio Vol Targeting

Disabled by default (vol_target_enabled=0). When enabled, all position sizes are uniformly scaled by target_ann_vol / forecast_vol (never above 1.0). Default target: target_ann_vol=0.12 (12% annualized). This keeps realized portfolio volatility consistent across different market environments.

Buying Power Guard

Before every order, the system checks live buying power from Alpaca with a 5% safety margin. If the order would exceed available buying power, it's automatically downsized (never blocked entirely).

Kelly Capital Sizing

The system can dynamically rebalance the HMM/Brando capital split each quarter using Kelly sizing based on actual closed-trade performance. Disabled by default (kelly_rebalance_enabled=0).

Kelly Formula

• Lookback: last 60 closed Brando trades
• Minimum trades required: 20 (falls back to static 60/40 split otherwise)
• Wins/losses measured by R-multiple: realized_pnl / risk_notional
• Negative Kelly expectancy → Brando allocation set to 0% (strategy disabled)
• Rebalance fires quarterly (first trading day, 17:00 ET)

Pool Management & Symbol Changes

The two active strategies handle pool changes differently. HMM is a rebalancer (reconciles positions to a target every cycle), while Brando is a position manager (exits follow pre-set rules, independent of the universe).

Max Positions

Each sleeve has an independent max positions limit, configurable from the Capital Allocation card on the Bot Engine page. For HMM this caps how many assets can be held simultaneously from the rotational pool. For Brando it caps the number of concurrent swing trades. Both accept a value between 1 and 50.

HMM Feature Engineering

The HMM classifies environments by observing Z-score standardized technical indicators (252-bar rolling lookback). Features are split into Core and Extended sets.

Walk-Forward Backtest

The backtest uses a rolling window approach to simulate realistic out-of-sample performance. No future data is used at any point — the HMM is retrained on each in-sample window.

Window Design

• In-sample: 252 bars (1 year) — HMM is trained on this window
• Out-of-sample: 126 bars (6 months) — strategy is evaluated here
• Step: 126 bars — windows slide forward by half a year
• Fill delay: 1 bar (orders execute on the next bar's open)
• Rebalance threshold: 10% (only rebalances if allocation changes by >10%)

Regime-Aware Slippage

Baselines

• Buy-and-hold: 100% SPY, no rebalancing
• 60/40: 60% equity / 40% bonds (simulated 4% annual return)
• Vol-targeting: Inverse 20-day realized vol, targeting 10% annualized
• Dumb HMM: 3 fixed regimes, 1-day returns only, static allocation map

Metrics Computed

Total return, CAGR, annualized volatility, Sharpe ratio, Sortino ratio, maximum drawdown, and Calmar ratio — for both the strategy and all baselines across each OOS window.

Chart & Visualization Reference