MPF Market-Timing Study: A Historical Backtest of 2,464 Rules

Out-of-sample test: 160 timing versions, 2015–2026

Separate, additional study. This section describes a different study from the "2,464 parameter sets" study above. Their design, data period and rules differ, and their figures are not directly comparable.

Study design

• Rules: 28 timing rules plus continuous holding (buy-and-hold) as the baseline, 29 in total, in ten families:
  • trend (T1–T4, e.g. 10-month and 200-day moving averages)
  • momentum (M1–M2)
  • drawdown triggers (D1–D3)
  • volatility (V1–V2, VIX)
  • credit (C1–C2, high-yield vs Treasury fund ratios)
  • macro (X1, Sahm rule / yield-curve)
  • valuation (X2, CAPE)
  • exit-and-re-entry rules (E1–E6)
  • combinations (K1–K7)
• How the 160 versions arise:
  • 26 rules were each tested at three check frequencies: monthly (month-end), bi-weekly (15th and month-end) and daily.
  • T1 and K7 were tested monthly only.
  • That gives 28 + 26 + 26 = 80 versions per equity mix, and 160 across the two mixes.
• Equity mixes:
  • "v2 mix": global 45%, US 20%, Asia ex-Japan 12.5%, Hong Kong 15%, Japan 7.5%.
  • "Global mix": ACWI from April 2008; before that, 60% S&P 500 index fund plus 40% international index fund.
• Holding range:
  • Equity held between 70% and 100%, with the rest in an MPF Conservative / money-market proxy.
  • Signals are read at the close; money is out of the market for 3 days and reinvested 4 business days after the signal.
• Periods:
  • Rules designed on October 2002 to December 2014.
  • Tested out of sample on January 2015 to September 2026.
• Data:
  • Equity: daily adjusted closes of US-listed funds from Yahoo Finance (VFINX, VGTSX, EPP, EWH, EWJ, ACWI), less a 1% a year MPF-style fee.
  • Cash: August 2020 to August 2026, the actual average of 29 MPF Conservative Funds; 2012 to July 2020, 1-month HIBOR less 0.8%; earlier, the US 3-month Treasury bill less 0.8%.
  • Signal inputs also use VIX, high-yield and Treasury fund prices, BLS, US Treasury and multpl.com data.
  • The equity data are not actual MPF fund daily prices.
• Sources: timing/README.md; timing/scripts/study.py, metrics.py, timing_core.py; timing/out/signal_catalogue.csv (from audit_bundle.zip, SHA-256 verified).

Results (January 2015 to September 2026, out of sample)

• 147 of 160: 147 of the 160 versions had a lower annualised return than holding 100% equity throughout; 13 had a higher one and none tied.
  • Holding returned 8.99% a year on the v2 mix and 9.85% on the global mix.
  • By mix: 75 of 80 v2-mix versions and 72 of 80 global-mix versions were lower.
  • Source: signal_catalogue.csv (column bnd_cagr_OOS 2015-26).
• Distribution: annualised returns across the 160 versions:
  • lowest 6.64%, highest 10.46%;
  • 25th percentile 8.12%, median 8.73%, 75th percentile 9.25%.
  • The median gap against the matching buy-and-hold was −0.64 percentage points a year.
  • Same source and period (2015-01..2026-09).
• Check frequency and switching (v2 mix):
  • Monthly checks: 28 versions, a median of about 1.4 switches a year (1.36), median return 8.59%.
  • Bi-weekly checks: 26 versions, about 1.8 a year (1.83), 8.44%.
  • Daily checks: 26 versions, about 5.9 a year (5.87), 7.97%.
  • On the global mix the medians were 9.35%, 9.39% and 8.89%.
  • Across both mixes, all 52 daily-checked versions returned less than holding.
  • Source: signal_catalogue.csv (bnd_cagr_OOS 2015-26, bnd_switches_yr_OOS 2015-26, grouped by freq); 2015-01..2026-09.
• Drawdowns:
  • 139 of the 160 versions had a shallower month-end maximum drawdown than holding.
  • The median was −21.43%, against −23.89% for holding the v2 mix and −26.29% for the global mix.
  • Most versions gave up return for a shallower fall.
  • Source: signal_catalogue.csv (bnd_max_dd_OOS 2015-26); 2015-01..2026-09.
• Design period for comparison:
  • In October 2002 to December 2014, 92 of the 160 versions beat holding.
  • The Spearman rank correlation between design-period and test-period returns was 0.06 for the v2 mix and −0.04 for the global mix.
  • Versions that did better in the design period showed no tendency to keep leading.
  • Source: signal_catalogue.csv (bnd_cagr_IS 2002-14).
• These are historical results: most out-of-sample versions underperformed continuous holding, while a minority (13) outperformed over the same period.

Caveats

1. A longer period gives a different answer:
   • This is a walk-forward run of the same study as this section (the same 28 rules and settings). It is not the 2,464-set study above.
   • Method:
     • Each January, pick the version with the highest cumulative return from October 2002 to the previous December, and hold it for 12 months.
     • The candidates are the monthly- and bi-weekly-checked versions only: 54 per mix, excluding the daily-checked ones.
     • Same 70–100% equity band, fees and 4-business-day switch delay.
   • Picks: v2 mix: K3 (monthly) every year 2006–2020, then K2 (bi-weekly) 2021–2026.
   • Results:
     • January 2006 to September 2026: 7.89% a year, slightly above the 7.61% for buy-and-hold (v2 mix). The month-end maximum drawdown was −43.3%, against −54.7% for holding.
     • On the global mix it was 7.65% vs 7.42%.
     • Over 2015–2026 alone, the same method returned 8.31%, below 8.99% for holding.
   • Caveat: the rules were designed on 2002–2014 data, so the 2006–2014 part is not out of sample.
   • A separate 7.61%: the global equity proxy benchmark in the 2,464-set study above (October 2006 to September 2026) is also 7.61%. It is a different series, and the two should not be conflated.
   • Sources: timing/out/walkforward.csv, walkforward_picks_v2mix.csv, walkforward_picks_global.csv; timing/scripts/evaluate.py (walk-forward block).
2. The test period was not fully blind:
   • The rules were designed by people who had lived through 2015–2026.
   • Some rules (e.g. K3) were chosen after the test-period results were visible.
   • Many rules and parameters were tried in total, which creates data-snooping bias.
   • Source: AUDIT_REPORT.md §3.9, §6.
3. Month-end prices understate drawdowns:
   • Returns and maximum drawdowns here are simulated daily but measured on month-end values.
   • In sharp shocks, month-end figures can be about 10–17 percentage points shallower than daily ones: COVID 2020, S&P 500 fund 14.2, ACWI 12.5, EWH 10.5 points (2800.HK 6.2); April 2025, 11.2–16.9 points.
   • For the worst drawdown of the whole 2015–2026 window, the v2 mix (fixed weights, before fees) differed by about 10.8 points (−31.9% daily vs −21.1% month-end); individual proxy funds differed by 2.2–12.5 points.
   • Source: timing/data/daily_adjclose.csv, our calculation.
4. Proxies:
   • The equity data are overseas-listed funds (in USD, assuming the HKD peg), not actual MPF fund daily prices.
   • Holdings, hedging, fees and pricing times differ from MPF funds.
5. A separate study: the "2,464 parameter sets" study above covers roughly 2006–2026 and tests "two-branch" and "equity core + tactical cash" rules. It is a different study from this section.
6. Past performance is not indicative of future results. This section describes historical backtests only. It is not a forecast, is not advice to buy, sell, hold or switch any fund, and refers to no individual fund.