Abstract

This paper presents a comprehensive analysis of three regime-based portfolio allocation strategies tested over different time horizons. We compare: (1) an original regime-detection strategy with Bitcoin (V1, 2019-2025), (2) an improved version with dynamic optimization (V2, 2019-2025), and (3) a long-term test without Bitcoin (2000-2025). Our findings challenge the conventional wisdom that increased complexity leads to superior returns.

Introduction

The quest for superior risk-adjusted returns has driven the development of increasingly sophisticated portfolio allocation strategies. Regime-based approaches, which adapt asset allocations based on detected market conditions, represent a natural evolution from static allocation methods. This paper examines three iterations of a regime-based portfolio strategy, each representing different levels of sophistication and tested over varying time horizons.

Motivation

This research addresses three critical questions:

1. Can machine learning-based regime detection improve portfolio performance?
2. What is the true cost of complexity in terms of transaction costs and overfitting?
3. Under what conditions do simple strategies outperform sophisticated alternatives?

The strategy defines three market regimes: Defensive (high volatility, risk-off), Neutral (transitional), and Aggressive (low volatility, risk-on). V1 used fixed VIX thresholds and MA filters; V2 added golden cross, multi-horizon momentum, and forward-looking signals.

Key Equations

Dynamic allocation optimization (V2) solves a Sharpe-maximizing problem within regime-specific bounds:

where and are estimated from a 60-day rolling lookback.

Total transaction cost per rebalance:

Performance metrics used throughout:

Algorithm Blueprint

Regime-Based Portfolio Allocation

Inputs: Daily returns data, asset universe (SPY, TLT, BTC, GLD), regime transition bounds

Outputs: Portfolio weight series, net returns after costs

Algorithm:

1. REGIME DETECTION (daily)

   • Compute VIX percentile, SPY price vs MA(60), yield curve slope
   • V1 Strategy: Aggressive if AND price AND yield curve not inverted
   • V2 Strategy: Aggressive if AND golden cross active AND
   • Defensive regime for opposite conditions, Neutral otherwise

2. PER-REGIME OPTIMIZATION

   • Estimate mean returns and covariance from rolling 60-day window
   • Solve maximum Sharpe ratio problem within regime-specific bounds:
   • Allocation bounds by regime:
     • Defensive: SPY [0-25%], TLT [20-50%], BTC [0-5%], GLD [15-40%]
     • Neutral: SPY [30-60%], TLT [15-35%], BTC [0-10%], GLD [10-25%]
     • Aggressive: SPY [50-75%], TLT [5-15%], BTC [5-20%], GLD [5-15%]

3. EXECUTION & COST ACCOUNTING

   • Portfolio turnover:
   • Net returns:
   • Transaction costs: 30bps round-trip

4. BENCHMARK COMPARISON

   • 60/40 benchmark:
   • Compare all strategy versions against benchmark using Sharpe ratio, CAGR, and maximum drawdown

Strategy Flow

Results

| Version | Test Period | Assets | Aggressive % | CAGR | Sharpe | Verdict | |---|---|---|---|---|---|---| | V1 (Original) | 2019-2025 | With BTC | 0.0% | +0.11% | -1.10 | Failed | | V2 (Improved) | 2019-2025 | With BTC | 13.1% | +0.21% | -0.88 | Marginal | | No BTC (Honest) | 2000-2025 | No BTC | 27.8% | -0.02% | -0.98 | Failed | | 60/40 Benchmark | 2000-2025 | — | — | +0.43% | -0.70 | Winner |

V1 never triggered the aggressive regime (0.0%), missing the 2025 bull market entirely. The 2% annual transaction cost differential (AI strategy: ~2.5%/yr vs 60/40: ~0.3%/yr) compounds to approximately -35% over 21 years.

Quantified bias sources inflate backtest results by an estimated 4–9% over the test period, with look-ahead bias (+1–2% annual) and survivorship bias from Bitcoin (+0.5–1.5% annual) as the largest contributors.

Contributions

• Comprehensive three-way comparison of regime-switching strategies against a static benchmark with full transaction cost accounting
• Honest quantification of 10 bias sources that inflate backtest performance, totalling +4% to +9.3% over the test period
• Empirical demonstration that the 60/40 portfolio is a hard-to-beat baseline in realistic deployment conditions
• Practical guidelines: apply 20–40% haircut to reported backtest returns; require 20+ years of out-of-sample data before deployment

Paper

Author

Frankline Misango Oyolo Quantitative Research Division, Arithmax Research Frankline@arithmax.com — Published: March 5, 2026