Skip to content
Tutorials Center
My learning

Topics

Lesson 17 of 40 ยท Smoothing and Exponential Methods

Holt's Linear Trend Method

Extend exponential smoothing to trending data with Holt's method: level and slope equations, the beta parameter, damped trends that stop runaway forecasts, and statsmodels examples.

Simple exponential smoothing predicts a flat line, so on a series with a trend its forecasts fall behind. Holt's linear trend method (also called double exponential smoothing) fixes this by tracking two things: a level and a slope. Forecasts then continue along a line. This lesson explains the equations, the extra smoothing parameter, and the damped trend variant that is often the safest choice for real forecasting.

Two equations, two states

Holt's method updates a level l and a trend (slope) b, each with its own smoothing parameter:

  • Level: l(t) = alpha * y(t) + (1 - alpha) * (l(t-1) + b(t-1))
  • Trend: b(t) = beta * (l(t) - l(t-1)) + (1 - beta) * b(t-1)
  • Forecast: yhat(t+h) = l(t) + h * b(t)

The level is the smoothed value, the trend is the smoothed per-step change. A large beta lets the slope change quickly; a small beta keeps it stable. Both lie between 0 and 1.

Python Holt's recursion by hand

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The slope settles near the average step of about 2 units per period, and the forecasts keep climbing from the last level.

Fitting with statsmodels

The Holt class estimates alpha, beta and the initial values:

Python Holt vs SES on a trending series

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Holt's forecast error is much smaller because its forecast rises with the data. The estimated beta is typically small for steady trends: the slope barely changes, so it should be averaged over a long history.

The danger of straight-line forecasts

Holt's trend continues forever. Over a long horizon this over-forecasts when real growth slows down (markets saturate, populations level off). Hyndman and Athanasopoulos demonstrated that for many series the linear trend is too optimistic.

Damped trend

A damped trend multiplies the slope by a factor phi (between 0 and 1, typically 0.8 to 0.98) at every future step, so the trend flattens out:

yhat(t+h) = l(t) + (phi + phi^2 + ... + phi^h) * b(t)

As h grows the sum approaches phi / (1 - phi), so the forecast converges to a constant. With phi = 1 you recover Holt's plain method.

Python Damped versus linear trend forecasts

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

When growth is slowing, the damped forecast stays close to the actual values while the linear one keeps climbing too steeply. In the large forecasting competitions damped trend methods were among the most reliable defaults, which is why many automatic systems prefer them.

Initial values and options

  • initialization_method="estimated" lets statsmodels optimise the starting level and slope; "heuristic" and "legacy-heuristic" use simple rules; "known" lets you supply them.
  • exponential=True gives a multiplicative trend (growth by a constant percentage rather than a constant amount).
  • You may fix parameters: .fit(smoothing_level=0.8, smoothing_trend=0.2, optimized=False).
Python Linear versus exponential (multiplicative) trend

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

For compound growth the multiplicative trend matches almost exactly, while the straight line falls short.

A practical recipe

  1. Plot the data. Is there a trend? If not, use SES.
  2. Fit Holt, damped Holt and possibly multiplicative Holt.
  3. Compare them with a time-series cross-validation, not just in-sample error.
  4. Prefer the damped variant when the horizon is long or the trend might fade.
  5. If the data are seasonal, deseasonalise or use Holt-Winters.

Check your understanding

Quick check ยท 4 questions

of 4 answered You scored / 4 ยท perfect!

  1. 1. What states does Holt's method track?

    Holt maintains a smoothed level and a smoothed trend (slope), each with its own parameter.
  2. 2. What does the damping parameter phi do?

    Each step ahead multiplies the slope contribution by phi, so the forecast converges to a constant.
  3. 3. Growth is about 2 percent per period, compounding. Which option suits best?

    A multiplicative trend grows by a constant factor per step instead of a constant amount.
  4. 4. You forecast 5 years ahead and fear growth will slow. Which method is a good default?

    Damping prevents the trend from running away over long horizons.

Summary

  • Holt's method smooths a level and a slope and forecasts a straight line.
  • alpha controls the level, beta the slope; both are estimated by statsmodels.
  • A damped trend (damped_trend=True) flattens the forecast and is a robust default for long horizons.
  • exponential=True gives a multiplicative trend for percentage growth.
  • Holt ignores seasonality, which is the job of the next method.

Next: The Holt-Winters Seasonal Method