IS6400 Chap.6 Holt Trend Forecasting
Holt Trend Forecasting
Two states describe a moving series
Holt's method maintains a local level and a local trend. Each new observation updates the level using the previous level advanced by the old trend, then updates trend using the new movement in level. A forecast h periods ahead equals the latest level plus h copies of the latest trend.
Evaluation must move forward in time
Random folds destroy the information boundary.
Roll the forecast origin through history and evaluate the same horizon used by the operational decision. Compare a naive baseline, inspect signed error and report whether shocks or seasonal patterns leave systematic residual structure.
Keep period labels attached to every state
Write the old level and trend before inserting the new observation.
The level update blends that observation with the previous projected path; the trend update blends the new change in level with the old trend. Updating both at once invites an index error.
Afterward, verify that forecasts at successive horizons differ by exactly the new trend and that a negative trend lowers rather than raises later forecasts.
Match the backtest to the calendar
If orders are placed four weeks ahead, rolling one-step forecasts do not test the decision. Move an origin through history, use only information available at each origin, and predict four weeks ahead every time.
Store origin, horizon, prediction, actual and signed error in an audit table.
Compare with a naive last-value or drift baseline on exactly the same origins, including promotions, closures and shocks rather than quietly excluding difficult periods.
Read residual patterns as model boundaries
Persistent one-sided error signals bias, repeating calendar error signals missing seasonality, and a jump after a structural event may require a reset or explanatory variable.
Extreme smoothing weights can imitate recent history without solving those omissions. Report error by horizon and regime, connect the direction of misses to inventory or staffing cost, and state where uncertainty becomes too wide for the normal action. Sometimes escalation or a buffer rule is safer than a sharper point forecast.
What this chapter covers
- 01
Chronological data splits
- 02
Local level state
- 03
Local trend state
- 04
Smoothing weights
- 05
Forecast horizon
- 06
Rolling backtests and signed bias
Worked example · free
Update level and trend before forecasting
- 3Update the level. The mark allocation shown is not a University assessment scheme.
- 3Update the trend.
- 3Project the requested horizon and interpret direction.
Key terms
- Local Level
- The current baseline state around which a time series is estimated to move.
- Local Trend
- The current estimated change in level per period.
- Forecast Horizon
- The number of future periods between the forecast origin and the predicted outcome.
- Rolling Origin
- An evaluation design that repeatedly advances the training end and predicts later periods.
- Forecast Bias
- A tendency for errors to stay mainly positive or mainly negative rather than balancing around zero.
Holt Trend Forecasting FAQ
Why is a random train-test split unsafe for forecasting?
It allows later observations to influence a model judged on earlier ones and fails to mimic deployment. Chronological splits and rolling origins preserve the direction in which information becomes available.
How does the forecast change with horizon?
Holt's non-seasonal forecast adds the latest trend once for every step ahead. A positive trend raises later horizons; a negative trend lowers them. Error usually grows with horizon, so evaluation must match the operational lead time.
When is Holt's method not enough?
Repeating seasonal residuals, abrupt regime changes or nonlinear state-dependent behaviour indicate missing structure. Compare an expanded model with the same rolling protocol rather than forcing smoothing weights to absorb every pattern.
Exam move
Draw four boxes labelled old level, old trend, new observation and new states. Recompute one update while keeping period labels visible. Produce one-step and multi-step forecasts and confirm their difference equals the extra number of trend increments. Practise a rolling-origin sketch, then diagnose three residual patterns: random misses, sustained one-sided bias and repeated seasonal error.
Link each pattern to a different next investigation. Add a negative-trend example and check that later horizons fall by one trend increment each time. Build a five-row origin table with information cutoff, horizon, prediction, actual and signed error; compare the same rows with a naive baseline.
Mark promotions and shocks rather than deleting them, and state whether each miss suggests a reset, seasonal structure, an explanatory variable or an escalation outside the model's normal operating range. Finish by checking that every error uses the same units and horizon as the decision, then explain what a positive signed error means operationally.