Drift Forecasts And Linear Extrapolation

What a Drift Forecast Extends

A time series may rise or fall overall even when its observations fluctuate from one period to the next. A drift forecast summarizes that overall movement as an average change per period, then extends the same change into the future. The result is a straight-line forecast path anchored at the latest observation.

This method is useful when a series has a persistent overall direction and a simple extrapolation is required. It does not claim that every future observation will change by exactly the same amount. Rather, it applies a constant-change rule to produce a forecast. Whether that rule is sensible depends on the context and on whether the observed net movement is a useful guide to the future.

Definition: Let $y_1,y_2,\ldots,y_T$ be $T$ equally spaced observations, with $T\geq 2$. The drift estimate is the average change between the first and latest observations: $$ \hat{d}=\frac{y_T-y_1}{T-1}. $$ The forecast $h$ observation intervals beyond the latest value is $$ \hat{y}_{T+h}=y_T+h\hat{d},\qquad h=1,2,\ldots. $$ Here, $\hat{d}$ is measured in units of the series per observation interval, and $h$ is counted in those same intervals.

The denominator is the number of elapsed intervals, $T-1$, not the number of observations, $T$. For example, six monthly observations contain five month-to-month intervals. This distinction matters whenever the drift is calculated from data indexed by observation number.

Why the Formula Gives a Straight-Line Forecast

The average change $\hat{d}$ is the slope of the line joining the first observation to the latest one. Extending that line beyond the latest time point gives the drift forecast. To see the connection, label the observations by $t=1,\ldots,T$. The line through $(1,y_1)$ and $(T,y_T)$ has slope $(y_T-y_1)/(T-1)$. At time $T+h$, its value is its value at $T$, plus $h$ times the slope.

\begin{aligned} \text{slope} &=\frac{y_T-y_1}{T-1}=\hat{d},\\ \text{line value at }T+h &=y_T+\big((T+h)-T\big)\hat{d}\\ &=y_T+h\hat{d}. \end{aligned}

Thus the forecast is the latest observed value plus the estimated average change multiplied by the number of future intervals. If the drift is positive, the forecast rises linearly; if it is negative, the forecast falls linearly; and if the first and latest observations are equal, the drift is zero and the forecasts remain at the latest value.

The method uses the two endpoints to estimate its slope. Intermediate observations do not enter the drift calculation, although they may be important when judging whether a straight-line extrapolation is reasonable. This is not the same as fitting a least-squares trend line, which uses all observations to determine a slope and intercept.

Calculating Forecasts in the Series’ Time Units

Before calculating, identify the observation spacing and the forecast horizon. If the data are monthly, $h=3$ means three months ahead. If the data are quarterly, $h=3$ means three quarters ahead. For a horizon stated in a different unit, convert it into the number of observation intervals before applying the formula.

Worked Example: Extending a Quarterly Increase

A hypothetical insurer tracks a quarterly service-volume index. The six quarterly observations are 42, 45, 48, 51, 54, and 57 index points. Use the drift method to forecast one quarter and four quarters beyond the latest observation. The series is treated deterministically for this calculation: no probability distribution, interest rate, or mortality assumption is used.

Step 1: Identify the endpoints and intervals. There are $T=6$ observations. The first value is $y_1=42$, and the latest value is $y_T=y_6=57$. Six quarterly observations span $T-1=5$ quarterly intervals.

Step 2: Calculate drift per quarter.

\hat{d}=\frac{57-42}{6-1}=\frac{15}{5}=3 \quad\text{index points per quarter}.

Step 3: Forecast one quarter ahead. Set $h=1$ and use the latest observed value as the starting point:

\hat{y}_{7}=57+1(3)=60 \quad\text{index points}.

Step 4: Forecast four quarters ahead. Four quarters are four observation intervals, so $h=4$:

\hat{y}_{10}=57+4(3)=57+12=69 \quad\text{index points}.

Check. The forecast path adds 3 points per quarter: 60, 63, 66, and 69 over the next four quarters. The fourth-quarter forecast is therefore 12 points above 57, as the formula gives. The method avoids assuming that the forecast horizon of four quarters means four observations in the historical sample; it means four intervals beyond the latest observation.

Worked Example: Intermediate Fluctuations and Endpoint Drift

A hypothetical claims team records a monthly workload index over seven consecutive months: 240, 247, 251, 263, 268, 272, and 276. Calculate forecasts two and four months beyond the final observation. Treat the index as a deterministic series; no probability distribution, interest rate, or mortality assumption is specified or needed.

Step 1: Determine the time span. There are seven observations, so there are six monthly intervals between the first and the latest. The endpoints are $y_1=240$ and $y_7=276$.

Step 2: Estimate the monthly drift.

\hat{d}=\frac{276-240}{7-1}=\frac{36}{6}=6 \quad\text{index points per month}.

The observed increases between successive months are not all 6 points. The drift uses the total endpoint increase of 36 points spread over six intervals; it does not average the intermediate changes separately for a different result.

Step 3: Forecast two months ahead. With $h=2$:

\hat{y}_{7+2}=276+2(6)=276+12=288.

Step 4: Forecast four months ahead. With $h=4$:

\hat{y}_{7+4}=276+4(6)=276+24=300.

Check. Starting at 276, the forecast path increases by 6 points each month. After two intervals it reaches 288, and after four it reaches 300. These are extrapolations from the endpoints, not forecasts that reproduce the historical month-to-month fluctuations.

Negative Drift and Converting a Horizon

Drift can be negative when the latest observation is below the first. The arithmetic remains the same: a negative $\hat{d}$ is multiplied by the positive number of future intervals and added to the latest value. A horizon stated in months, weeks, or years must first be translated to the intervals used by the series.

Worked Example: A Declining Quarterly Measure and a Six-Month Horizon

A hypothetical benefit-administration team tracks a quarterly processing measure, in units, over five quarters: 60, 56, 51, 47, and 40. Forecast six months and fifteen months beyond the latest observation. Since observations are quarterly, take one quarter to be three months. The calculation is deterministic; it uses no probability distribution, interest rate, or mortality assumption.

Step 1: Calculate the quarterly drift. Five observations span four quarters. The first value is 60 and the latest is 40:

\hat{d}=\frac{40-60}{5-1}=\frac{-20}{4}=-5 \quad\text{units per quarter}.

Step 2: Convert six months to quarters. At three months per quarter, six months is two quarters, so $h=2$:

\hat{y}_{T+2}=40+2(-5)=40-10=30.

Step 3: Convert fifteen months to quarters. Fifteen months is five quarters, so $h=5$:

\hat{y}_{T+5}=40+5(-5)=40-25=15.

Check. The line declines by 5 units per quarter: 35 after one quarter, 30 after two, and 15 after five. The negative sign is preserved in each calculation, so the forecasts are below the latest observation of 40.

Assumptions, Interpretation, and Common Traps

A drift forecast requires a numeric series with equally spaced observations and at least two values. Its calculation imposes a constant average change per observation interval beyond the latest value. It does not require the historical observations to increase or decrease by identical amounts, and it does not establish that the future will follow the extrapolated line. The resulting values should be interpreted as forecasts under the drift rule, not as certainty about future outcomes.

Because the estimate uses only the first and latest observations, a large endpoint movement can dominate the result even if the intermediate data show a different pattern. Before relying on the forecast, inspect the series in context. A straight-line extension may be unhelpful if the endpoints are atypical, the direction is changing, or a known limit makes continued linear movement implausible. The formula itself does not adjust for seasonality or other systematic patterns.

Common exam trap: Dividing the endpoint change by $T$ instead of $T-1$ treats the number of observations as the number of elapsed intervals. A second common error is to use a horizon in calendar units directly as $h$, even when it differs from the series’ observation interval. Count the elapsed intervals first, then convert the forecast horizon to the same units.

For example, in the quarterly decline above, an incorrect calculation might divide the endpoint change by five observations:

\text{Wrong: }\frac{40-60}{5}=-4 \quad\text{units per quarter}.

That denominator is wrong because five quarterly observations span only four quarterly intervals. The correct drift is $-20/4=-5$ units per quarter. Also, six months is two quarters, not six quarters. The correct six-month forecast is $40+2(-5)=30$; using $h=6$ would incorrectly forecast $40+6(-5)=10$, a fifteen-month rather than six-month horizon.

A useful calculation check is to state the units at each stage: endpoint change in series units, drift in series units per interval, and forecast horizon in intervals. Then $h\hat{d}$ has the same units as the observations and can be added to $y_T$. If those units do not match, revisit the interval count or conversion.

Check Your Understanding

For each problem, calculate the endpoint-based drift and apply it using the stated observation interval.

A monthly index has eight consecutive observations. Its first value is 105 and its latest value is 126. What is the drift and the forecast five months ahead? (Answer: drift $=3$ index points per month; forecast $=141$.)
A weekly measure has nine consecutive observations, from 52 at the start to 68 at the end. What is its drift and forecast six weeks ahead? (Answer: drift $=2$ units per week; forecast $=80$.)
A quarterly measure has five observations, beginning at 95 and ending at 71. What is the drift and forecast three quarters ahead? (Answer: drift $=-6$ units per quarter; forecast $=53$.)
A weekly workload index has 13 consecutive observations, with first value 400 and latest value 460. Forecast eight weeks ahead. (Answer: drift $=5$ index points per week; forecast $=500$.)
A hypothetical series is observed every two weeks. There are six consecutive observations, with first value 30 and latest value 50. Treat three months as six weeks. What is the drift and forecast three months ahead? (Answer: drift $=4$ units per two-week interval; forecast $=62$, using $h=3$ intervals.)

What You'll Learn