Dynamic Mean Optimization

Static thresholds work when the underlying indicator has a stable typical level — but most indicators don't. RSI tends to oscillate around 50 in calm markets and around 60 in strong uptrends. A 30/70 entry/exit threshold is the right rule for one of those regimes, the wrong rule for the other. Dynamic Mean Optimization solves this by letting thresholds track the indicator's rolling mean with calibrated offsets — adapting automatically as the indicator's typical level shifts.

The Core Idea

Instead of "enter long when RSI < 30," the dynamic-mean rule is:

The threshold is no longer a fixed number. It tracks a rolling average of the indicator and sits a calibrated distance below it. When the indicator's typical level drifts up (e.g., RSI averaging 60 in an uptrend), the entry threshold drifts up too. The strategy fires when the indicator deviates from recent typical behavior, not from a static reference.

The Two Tuning Parameters

Mean Window N

How many bars of history are used to compute the rolling mean. Short windows (10–30 bars) make the threshold adapt quickly but track noise. Long windows (100+ bars) make the threshold smooth but slow to adjust to genuine regime changes. Typical sweet spot is in the 30–60 range for daily-bar strategies.

Offsets (Upper and Lower)

How far above and below the rolling mean the entry/exit thresholds sit. Symmetric offsets are the simplest specification (e.g., enter long at mean − 15, exit at mean + 15). Asymmetric offsets capture asymmetric mean-reversion (e.g., enter at mean − 20, exit at mean + 10).

Hurst-Based Calibration

The offset that maximizes return per unit time depends on the speed of mean reversion. The Hurst exponent measures that speed:

• Hurst near 0.5 (random walk): the indicator doesn't systematically revert. Dynamic-mean thresholds add no value over static thresholds.
• Hurst < 0.5 (mean-reverting): dynamic-mean is appropriate. Stronger mean reversion (lower Hurst) supports tighter offsets — the indicator returns to mean quickly, so wider offsets just delay entry.
• Hurst > 0.5 (trending): mean reversion isn't working. Don't use dynamic-mean (or any mean-reversion approach) on trending series.

RG001RGMO computes the Hurst exponent automatically and uses it to calibrate the offset starting point. The user can then sweep around that starting point to find the actual optimum.

What Dynamic Mean Adds Over Static

• Automatic adaptation to indicator drift. No need to manually re-tune thresholds when the market regime changes.
• Better behavior in transitional periods. Static thresholds either fire too often (during expansion) or too rarely (during contraction). Dynamic thresholds maintain consistent firing frequency.
• Smaller parameter count. Surprisingly, dynamic-mean often uses fewer effective parameters than per-regime optimization while capturing similar regime-adaptation behavior.

What It Doesn't Solve

• It still assumes mean-reversion. If the indicator is genuinely trending (Hurst > 0.5), no threshold scheme — static or dynamic — will produce sustainable returns.
• It can't replace a regime classifier. Dynamic-mean adapts the threshold reference; it doesn't turn the strategy off when the regime is hostile. Combine with regime filters when needed.
• The mean window is itself a parameter. Sweeping over (mean window, offset) creates a 2D optimization problem that can still overfit.

The Honest Workflow

1. Verify mean-reversion with Hurst. Hurst < 0.5 is the prerequisite. Without it, dynamic-mean isn't the right approach.
2. Compare to static as baseline. Run static grid search first. The static version's composite OOS Sharpe is the bar that dynamic-mean must clear.
3. Sweep mean window and offset. 2D grid: window length on one axis, offset on the other. Look for plateaus.
4. DSR-correct the result. The 2D grid gives more cells than 1D static — more multiple-testing exposure.
5. Walk-forward both static and dynamic-mean. If dynamic-mean's composite OOS Sharpe doesn't materially exceed static's, the additional complexity isn't earning its keep.

The Bottom Line

Dynamic Mean Optimization is the right approach when (a) your indicator has drifting typical levels rather than a fixed reference, and (b) the underlying behavior is mean-reverting per Hurst. When both conditions hold, dynamic-mean often produces materially better walk-forward performance than static thresholds — and often comparable performance to per-regime optimization with fewer effective parameters. The discipline is the same as for any optimization: walk-forward validate, compare to a simpler baseline, and prefer the simpler method when the more complex one doesn't clearly win.