QuanTAlib/lib/trends_IIR/holt/Holt.md

HOLT: Holt Exponential Moving Average

Single smoothing tracks level. Double smoothing tracks trend. The elegance is not in complexity but in the admission that yesterday's direction matters.

Property	Value
Category	Trend (IIR MA)
Inputs	Source (close)
Parameters	period, gamma (default 0)
Outputs	Single series (HOLT)
Output range	Tracks input
Warmup	period bars
PineScript	holt.pine
Signature	holt_signature

• Holt's exponential smoothing extends simple exponential smoothing (EMA) by adding a second equation that explicitly tracks the local trend.
• Similar: DEMA, KAMA | Complementary: Forecast accuracy metrics | Trading note: Holt exponential smoothing; level + trend components for forecasting.
• Validated against TA-Lib, Skender, and Tulip reference implementations where available.

Overview

Holt's exponential smoothing extends simple exponential smoothing (EMA) by adding a second equation that explicitly tracks the local trend. The result is a 1-step-ahead forecast that adapts to both the level and direction of the time series. When applied to financial data, HOLT produces a trend-following line that anticipates price continuation rather than merely reacting to it.

Historical Context

Charles C. Holt published the method in 1957 at the Carnegie Institute of Technology, though the work remained an unpublished ONR report until 2004. The method was independently popularized by Peter Winters (who added seasonality, creating Holt-Winters), but the core two-equation system belongs to Holt. In forecasting literature, this is "double exponential smoothing" — not to be confused with DEMA (which is a different construct using two cascaded EMAs with lag compensation).

The crucial difference from DEMA: Holt explicitly decomposes the signal into level and trend components, then recombines them for forecasting. DEMA applies algebraic lag correction without explicit trend modeling.

Mathematical Foundation

Level Equation

L_t = \alpha \cdot y_t + (1 - \alpha) \cdot (L_{t-1} + B_{t-1})

The level L_t is a weighted average of the current observation and the previous level-plus-trend forecast.

Trend Equation

B_t = \gamma \cdot (L_t - L_{t-1}) + (1 - \gamma) \cdot B_{t-1}

The trend B_t is a weighted average of the observed level change and the previous trend estimate.

Output (1-Step-Ahead Forecast)

\text{HOLT}_t = L_t + B_t

Parameter Mapping

Parameter	Formula	Default	Range
Alpha (\alpha)	2 / (\text{period} + 1)	period=10 → 0.1818	(0, 1)
Gamma (\gamma)	User-specified or \alpha	0 (auto = \alpha)	[0, 1]

Special Cases

• \gamma = 0 (auto): Uses \gamma = \alpha, providing balanced level/trend tracking
• \gamma \to 0 (manual): Trend component freezes; degenerates toward pure EMA
• \gamma \to 1: Trend reacts instantly to level changes; high noise sensitivity

Initialization

• Bar 1: L_1 = y_1, B_1 = 0, output = y_1
• Subsequent bars use the full equations above

Interpretation

• Trend following: When HOLT is above/below price, the trend component provides a directional bias
• Crossover signals: HOLT crossing price suggests trend reversal
• Lead indicator: Unlike EMA, HOLT anticipates continuation via the trend term

Parameters

Parameter	Type	Default	Description
period	int	10	Lookback period for alpha calculation
gamma	double	0	Trend smoothing factor; 0 = auto (uses alpha)

Performance Profile

Operation Count (Streaming Mode)

HOLT(N, γ) tracks both level and trend via two EMA-like updates per bar. The dominant cost is two FMAs (level update and trend update) plus a final addition.

Operation	Count	Cost (cycles)	Subtotal
Level: FMA(α, src, decay×(level+trend))	1	4	~4
Level delta: new_level − prev_level	1	1	~1
Trend: FMA(γ, delta, gammaDecay×trend)	1	4	~4
Output: level + trend	1	1	~1
Total	4	—	~10 cycles

O(1) per bar. One of the fastest trends_IIR indicators — only 4 operations per bar. When γ = 0 (γ defaults to α), the trend EMA degenerates to a standard EMA with no trend correction. WarmupPeriod = N.

Batch Mode (SIMD Analysis)

Operation	Vectorizable?	Notes
Level update	No	Recursive: depends on previous level + trend
Trend update	No	Recursive: depends on previous trend and new level
Output (level + trend)	Yes	VADDPD once level and trend series complete

Both state variables are recursive. Batch mode provides no SIMD opportunity beyond the final addition. Holt's method is strictly serial.

Metric	Value
Time complexity	O(1) per bar
Space complexity	O(1) — level + trend only
Allocations	Zero in Update hot path
FMA usage	Level and trend equations
SIMD potential	Limited (serial dependency)
Warmup period	Same as period parameter

Limitations

1. Trend overshoot: In ranging markets, the trend component causes systematic bias (output overshoots actual price during reversals)
2. Gamma sensitivity: Small gamma changes dramatically alter behavior; requires careful tuning
3. No mean reversion: The additive trend model assumes perpetual directional movement
4. Initialization sensitivity: First-bar seeding (level=price, trend=0) means early outputs are biased
5. Not a filter: Unlike Butterworth or SSF, Holt has no defined frequency response — it is a forecasting model applied as a filter

References

• Holt, C. C. (1957). "Forecasting Seasonals and Trends by Exponentially Weighted Moving Averages." ONR Research Memorandum No. 52, Carnegie Institute of Technology.
• Holt, C. C. (2004). "Forecasting Seasonals and Trends by Exponentially Weighted Moving Averages." International Journal of Forecasting, 20(1), 5–10. (Republication of the 1957 report)
• Gardner, E. S. (1985). "Exponential Smoothing: The State of the Art." Journal of Forecasting, 4(1), 1–28.
• Hyndman, R. J. & Athanasopoulos, G. (2021). "Forecasting: Principles and Practice." 3rd ed., OTexts.

See Also

• EMA — Single exponential smoothing (level only)
• DEMA — Double EMA with algebraic lag correction (different approach)
• TEMA — Triple EMA cascade