Holt's linear trend forecasts a startup's sign-ups 24 months out at an implausible level. What do you change?

## Why the straight line blows up Holt's linear trend method carries two states, a level and a slope, and updates both each period: ``` level_t = alpha * y_t + (1 - alpha) * (level_{t-1} + trend_{t-1}) trend_t = beta * (level_t - level_{t-1}) + (1 - beta) * trend_{t-1} forecast_{t+h} = level_t + h * trend_t ``` The forecast function is a straight line in h with no upper bound. That is fine at h = 1 and often fine at h = 3. At h = 24 it is an extrapolation claim: it asserts that the growth rate observed over the recent past continues undiminished for two more years. For a startup's monthly sign-ups this is exactly the wrong claim. Early growth is fed by the easiest-to-reach users, referrals from a small enthusiastic base, and one-off launch effects. The fitted slope encodes that phase. Projected 24 steps forward it produces a figure that may exceed the plausible addressable market — the classic symptom that the model form, not the data, is wrong. ## The damped-trend fix Gardner and McKenzie's damped-trend method introduces a fourth parameter, phi, and applies it to the trend both in the state update and in the forecast: ``` level_t = alpha * y_t + (1 - alpha) * (level_{t-1} + phi * trend_{t-1}) trend_t = beta * (level_t - level_{t-1}) + (1 - beta) * phi * trend_{t-1} forecast_{t+h} = level_t + (phi + phi^2 + ... + phi^h) * trend_t ``` With `0 < phi < 1` the multiplier on the trend is a geometric sum rather than h, so each further step ahead adds less than the one before. Summing the geometric series gives the long-run limit: ``` forecast_{t+h} -> level_t + phi * trend_t / (1 - phi) as h grows ``` The forecast still respects the direction and short-run strength of the trend — at h = 1 the multiplier is just phi, close to 1 — but it flattens out to a finite ceiling instead of running away. At phi = 1 the geometric sum becomes h and the method reduces exactly to Holt's linear trend, which is why phi = 1 is the boundary case rather than a separate model. ## Choosing phi Phi is estimated with the other parameters by minimising squared one-step errors or maximising the state-space likelihood. In practice it is constrained to a band such as 0.8 to 0.98, for two reasons: - Below roughly 0.8 the trend is extinguished within a handful of steps, so the model is barely distinguishable from a level-only model and the fitted slope stops meaning anything. - Above roughly 0.98 the damping is so weak that over any horizon you actually forecast the model behaves like the undamped one, and the estimate becomes numerically hard to pin down. A fitted phi near the lower bound is itself informative: it says the recent slope carries very little information about the future, which is a finding worth reporting rather than hiding. ## Why not other fixes? **Lower beta.** Shrinking beta makes the slope estimate steadier by weighting recent changes less, but the forecast function is still `level + h * trend` — still linear, still unbounded. A smaller beta changes *which* line you extrapolate, not the fact that you extrapolate a line. **Clip the output.** Capping the 24-month forecast at some business-plausible ceiling produces a number no model generated, invalidates the prediction intervals, and hides the modelling error rather than fixing it. If you genuinely have an external ceiling — a total addressable market — that belongs in a saturating growth model, stated as such. **More training data.** More history helps estimate parameters, but a longer sample of a fast-growing phase will still produce a steep slope and the same unbounded extrapolation. The problem is the shape of the forecast function. ## When is an undamped linear trend defensible? When the horizon is short relative to the sample, when there is an external reason to believe the slope persists (a contracted ramp, a physically driven quantity), or when the two forms are compared within the ETS family and the undamped one wins on a corrected information criterion. Even then, saying out loud what the h-step forecast implies in business terms is the cheapest sanity check available: if the number is one the company could not physically serve, the model has told you something about itself, not about the future. ## The empirical argument Damped-trend exponential smoothing earned its reputation in the published forecasting competitions. In the M3 competition the damped-trend form was among the most accurate methods across a large and varied collection of series, and a headline conclusion of that work was that statistically sophisticated methods did not systematically beat simple ones. That is why damped trend is a sensible default for automatic forecasting over many series rather than an exotic option: it protects against the single most damaging failure mode of trended smoothing, which is a confident straight line into a horizon nobody believes.