Rolling Window Backtest Validation Study

Consultant-Level Technical Report

Date: 2026-02-19 Classification: Research — For Professional Review Version: 1.0

Executive Summary

This study tested the repeatability of grid trading portfolio returns across 3 quarterly rolling windows spanning approximately 30 months of cryptocurrency market data. Using a blind forward-test methodology (Year 1 parameters applied to Year 2 data), we evaluated two strategies, three portfolio allocation methods, and the predictive power of our asset rating system.

Key findings:

• Dynamic Grid MVO portfolios returned 54.9% mean (range: 13.9% to 91.2%)
• Basic Grid MVO portfolios returned 14.5% mean (range: 12.8% to 17.6%)
• Rating system correctly ranked top vs bottom half in 100% of windows (avg spread: 31.1pp)

1. Methodology

1.1 Rolling Window Design

  30 months of 4-hour candle data (≈5,400 candles per asset)
  
  ├── Window 1: Year 1 [months 1-12]   → Year 2 [months 13-24]
  ├── Window 2: Year 1 [months 4-15]   → Year 2 [months 16-27]
  └── Window 3: Year 1 [months 7-18]   → Year 2 [months 19-30]*
  
  * Partial window if data < 24 months from offset

Each window is a complete, self-contained experiment:

1. Year 1 — Parameter generation period. Grid upper/lower limits, entry price, and asset ratings are derived solely from this data.
2. Year 2 — Blind forward test. The grid bot trades using only Year 1 parameters. No Year 2 information is used in parameter selection.
3. Quarterly slide — Windows overlap by 9 months, providing partially independent observations.

Note on independence: Because windows share overlapping data, the observations are not fully independent. This reduces the effective sample size for statistical inference. We address this by reporting wide confidence intervals and avoiding overly precise claims.

1.2 Grid Trading Strategies

1.3 Asset Rating System

Each asset is rated on a composite score (0-100) comprising 5 available components:

• Performance (30%) — Grid return vs buy-and-hold, Calmar outperformance
• Risk-Adjusted (15%) — Sharpe ratio scoring
• Sortino (15%) — Downside risk-adjusted return
• Alpha (10%) — Excess return over buy-and-hold
• Liquidity (5%) — Set to 80/100 for all major exchange-listed tokens

Components not available in backtesting (sentiment, ETS, exchange suitability) are excluded and remaining weights renormalized. Grade thresholds: A (80+), B (65-79), C (50-64), D (<50).

1.4 Portfolio Allocation Methods

Three allocation approaches are compared, plus equal-weight as a baseline:

1. MVO (Maximum Sharpe) — Mean-variance optimization via gradient descent, maximizing the Sharpe ratio
2. Calmar CDaR — Gradient ascent on Calmar ratio (return/max drawdown), 40% max per asset
3. HRP (Hierarchical Risk Parity) — Ward's clustering + recursive bisection on the equity curve correlation matrix
4. Equal Weight — 1/N allocation as a naive benchmark

2. Data Summary

13 assets analyzed across 3 rolling windows.

3. Per-Window Portfolio Results

3.1 Dynamic Grid

* Partial window — Year 2 < 12 months

Dynamic Grid — Summary Statistics:

3.2 Basic Grid

* Partial window — Year 2 < 12 months

Basic Grid — Summary Statistics:

4. Strategy Comparison: Dynamic Grid vs Basic Grid

4.1 Head-to-Head by Window (MVO Sharpe)

Dynamic Grid won 3 of 3 windows.

4.2 Visual Comparison

  Portfolio Returns by Window (MVO Sharpe)
  ─────────────────────────────────────────
  W1    Basic  ▓▓▓▓ 13.2%
        Dynamic ████████████████████ 59.8%

  W2    Basic  ▓▓▓▓▓▓ 17.6%
        Dynamic ██████████████████████████████ 91.2%

  W3    Basic  ▓▓▓▓ 12.8%
        Dynamic █████ 13.9%

  Legend: ▓ = Basic Grid  █ = Dynamic Grid

5. Allocation Method Comparison

Dynamic Grid

  Mean Returns by Allocation Method (Dynamic Grid)
  ──────────────────────────────────────────────
  MVO Sharpe    ████████████████████████████████████ 54.9%
  Calmar CDaR   ████████████████████████████████████████ 61.2%
  HRP           ███████████ 16.4%
  Equal Wt      ████████████████████ 31.3%

Risk-Adjusted (Sharpe Ratio):

Basic Grid

  Mean Returns by Allocation Method (Basic Grid)
  ──────────────────────────────────────────────
  MVO Sharpe    ████████████████████████████████████████ 14.5%
  Calmar CDaR   ████████████████████████████████████ 12.9%
  HRP           ████████████████████████████████████ 13.2%
  Equal Wt      █████████████████████████████████████ 13.3%

Risk-Adjusted (Sharpe Ratio):

6. Rating System Validation

Dynamic Grid

Summary:

• Median-split hit rate: 100% (3/3 windows)
• Average top-half vs bottom-half spread: 31.1pp
• Average top-20% vs bottom-20% spread: 76.5pp
• Spearman rank correlation (score vs return): 0.668

Performance by Rating Tier (averaged across windows):

  Average Annual Return by Rating Grade
  ──────────────────────────────────────
  Grade A       ████████████████████████████████████████ 141.9%
  Grade B       ██████ 20.4%
  Grade C       ████ 13.6%
  Grade D      -░░ -7.7%

Basic Grid

Summary:

• Median-split hit rate: 100% (3/3 windows)
• Average top-half vs bottom-half spread: 5.6pp
• Average top-20% vs bottom-20% spread: 9.7pp
• Spearman rank correlation (score vs return): 0.478

Performance by Rating Tier (averaged across windows):

  Average Annual Return by Rating Grade
  ──────────────────────────────────────
  Grade B       ████████████████████████████████████████ 13.1%
  Grade C       ███████████████ 4.9%
  Grade D       ████████ 2.5%

7. Per-Asset Results (Dynamic Grid)

Mean Return Across All Windows:

  LTC           ████████████████████████████████████████ 101.4%
  XRP           █████████████████████ 52.0%
  ETH           ██████████ 25.5%
  LINK          ██████████ 25.1%
  BNB           █████ 12.4%
  ADA           ████ 10.7%
  BTC           ███ 7.7%
  SOL           ██ 3.9%
  UNI          -░ -3.4%
  POL          -░░░░ -11.1%
  DOT          -░░░░░░░ -18.4%
  AVAX         -░░░░░░░░ -19.8%
  DOGE         -░░░░░░░░░ -21.8%

8. Limitations and Caveats

1. Small sample size. 3 rolling windows with quarterly overlap yield approximately 1.5-2 effective independent observations. Statistical conclusions should be interpreted as directional, not definitive.
2. Overlapping windows. Windows share 9 months of data, creating autocorrelation in results. We report this transparently rather than claiming independence.
3. Survivorship bias. Only assets currently tracked are included. Tokens that were delisted or lost relevance during the study period are not represented.
4. Transaction costs. Simulations include estimated trading costs but real-world slippage, exchange fees, and withdrawal costs may differ.
5. Market regime. The study period (2023-2026) includes a significant crypto bull market. Results in a prolonged bear market may differ materially.
6. No leverage or margin. All simulations assume spot trading with no borrowed capital.
7. Partial windows. Any window with Year 2 < 12 months is flagged. Results are NOT annualized to avoid extrapolation bias.
8. Liquidity assumption. All 13 assets are assigned a liquidity score of 80/100 as a reasonable proxy for major exchange-listed tokens.

9. Conclusions

1. Returns are positive across all tested windows, with Dynamic Grid MVO ranging from 13.9% to 91.2%. However, the range is wide, and 3 windows is insufficient for narrow confidence bounds.
2. Dynamic Grid outperforms Basic Grid by approximately 40.4pp on average, primarily due to trailing-up level expansion in bull markets.
3. The rating system shows predictive signal (100% hit rate, positive rank correlation), but the small number of windows limits statistical significance. The directional evidence is encouraging.
4. MVO, Calmar, and HRP produce differentiated portfolios with varying risk/return profiles. No single method dominates across all windows, suggesting ensemble or rotation approaches merit further study.
5. Grid range coverage is the primary driver of individual asset returns. Assets where Year 2 price stayed within the Year 1-derived grid range performed dramatically better than those that escaped the range. This is the most important finding for strategy improvement.

Generated by Rolling Window Backtest Validation Study v1.0 13 assets, 3 windows, 2026-02-19T20:02:40.888Z