Forecasting is a constant challenge for Excel users. Business data doesn’t just grow in straight lines… it shifts with trends, cycles, and noise. Exponential smoothing is a family of methods that accounts for these patterns by weighting recent data more heavily while still learning from the past.
Excel does offer basic exponential smoothing through tools like the Analysis ToolPak and Forecast Sheet, but its features are limited compared to Python and Copilot, which make it easier to explore variations, see diagnostics, and understand what’s happening under the hood.
To show the full picture, we’ll use the classic Airline Passengers dataset, a monthly record from 1949–1960 with both a clear upward trend and repeating seasonal cycles, exactly the kind of structure exponential smoothing was built for.
You can follow along using the download file below:
Simple Exponential Smoothing
We’ll start simple with Simple Exponential Smoothing (SES). This is the most basic version of exponential smoothing. It assumes the data more or less bounces around a stable level with no trend and no seasonality.
Let’s see what happens if we apply SES to the airline passenger data:
“Fit a simple exponential smoothing (SES) model to the dataset and plot the smoothed line against the actual data. Use this to explain why SES alone isn’t enough when the data has a clear trend and seasonality.”
The forecast, shown by the orange dashed line, comes out completely flat. The model is basically saying, “passenger counts will just stay around the same level as the last year or so of data.” But the actuals, shown by the blue line, tell a different story. The series is clearly growing, and it has regular seasonal ups and downs.
This mismatch is important. SES did exactly what it was designed to do: smooth the data and project a constant level. The problem is that this dataset doesn’t fit those assumptions. Starting here gives us a baseline. Now we can ask, what happens if we use models that account for trend and seasonality?
Holt’s linear trend
Now let’s take things a step further with Holt’s linear trend method. This model adds a second smoothing component for trend, so instead of projecting a flat line like SES, it tries to extend the overall direction of the series into the future:
“Apply Holt’s linear trend method to the Airline Passengers dataset. Plot actual values through 1960, then overlay only the next 24 months of forecast. Explain how adding a smoothed trend improves the forecast compared to SES, but why it still doesn’t handle seasonality in the data.“
This time the forecast rises upward instead of staying flat. Compared to SES, that’s an improvement because Holt’s method is recognizing that the series has been trending up over time. But notice what’s missing. The actual passenger counts go up and down in a regular seasonal cycle, and the forecast doesn’t capture any of that. Holt’s method can follow the long-term direction of the data, but it still smooths away those repeating seasonal swings.
This is progress compared to SES, but we’re still not fully matching what the data is doing. Next we’ll see how adding in seasonality changes the picture.
Damped trend model
The next step is to look at a variation of Holt’s method called the damped trend model. Holt’s linear trend keeps extending upward at the same rate forever, which can be too aggressive for many real-world situations. The damped version adds a mechanism to gradually flatten the trend as the forecast goes further out.
“Fit a damped trend model to the Airline Passengers dataset. Plot the actual data through 1960, then add only the forecast for the next 24 months. How does this version of the trend look compared to Holt’s linear trend?“
In this forecast, the orange dashed line starts off like Holt’s linear trend but then slowly bends and flattens over time. Compared to the straight-line projection we saw earlier, this version looks more cautious. It avoids predicting unlimited growth at the same steep rate and instead tapers the trend.
This makes the damped trend especially useful when you think a series will keep growing but not as quickly as in the past. At the same time, it has the same limitation as Holt’s linear trend: it doesn’t capture the repeating seasonal ups and downs in the airline passenger data. That’s the next piece we’ll add.
Holt-Winters seasonal model
So far, we’ve tried SES, Holt’s linear trend, and Holt’s damped trend. Each one added something new, but none of them picked up on the repeating seasonal pattern that’s so obvious in the airline passenger data. The Holt-Winters model extends Holt’s method by adding a seasonal component.
“Fit a Holt-Winters seasonal model to the Airline Passengers dataset. Plot the actual values through 1960, then only the 24-month forecast. Do you see anything new in the way this forecast follows the patterns in the data?”
This time the forecast doesn’t just rise upward in a smooth curve. Instead, it rises and falls in a regular cycle, closely matching the seasonal peaks and dips we see in the actual data. That repeating wave pattern was missing from the earlier forecasts.
With Holt-Winters, we now have a model that reflects three things at once: the overall level of the series, the upward trend, and the repeating seasonal cycle. For this dataset, that combination makes the forecast much more realistic than what we saw with SES or Holt’s methods.
Forecast model comparison
So far we’ve been looking at the different smoothing models by eye. That’s helpful, but in practice we often want a more objective way to compare them. One common approach is to split the data into a training set and a test set, then measure how well each model’s forecasts match the actuals in the test period.
Fit SES, Holt’s linear trend, and Holt-Winters seasonal models. Split the data into training (up to 1958) and test (1959–1960). For each model, generate forecasts and compute accuracy metrics such as MAE, RMSE, and MAPE. Summarize the results in a comparison table and explain which model performed best.
The table shows accuracy metrics for each model on the 1959–1960 test period. SES has the largest errors, Holt’s linear trend does a bit better, and Holt-Winters clearly outperforms both. Its MAE, RMSE, and MAPE are all much lower than the other two models.
This makes sense given what we saw visually. The airline passenger data has an upward trend and a repeating seasonal cycle, so the model that can capture both ends up giving the best forecasts. SES was too simple and ignored both features. Holt’s method accounted for the trend but still missed the seasonal swings. Holt-Winters brought everything together.
This exercise shows the value of comparing models in a structured way. Instead of just relying on a chart, you can use error metrics to confirm which model is the best fit for your data.
Conclusion
Working through exponential smoothing step by step shows how each model adds something new. SES gives a flat forecast, which works only when the data has no structure. Holt’s method improves things by adding a trend, and the damped trend variation shows how you can keep that growth from running away. Holt-Winters takes it further by layering in seasonality, which is why it matched our particular dataset of airline passenger counts so well.
The next step in analysis, as usual, is broadening the toolkit. You might experiment with additive vs multiplicative seasonality, add forecast intervals to communicate risk, or move beyond exponential smoothing into ARIMA, SARIMA, or Prophet for more flexible modeling. Machine learning approaches like gradient boosting or neural networks can go even further when the data is complex. Many of these techniques can again be implemented directly in Excel with the help of Python and Copilot.
For Excel users, the key lesson is that internal tools like Forecast Sheet and the Analysis ToolPak is only one flavor of Holt-Winters behind the scenes. It’s a good starting point, but not the whole story. By learning the wider family of smoothing methods, you can see when Excel’s defaults fall short and when it’s worth exploring other tools. As with most analyst tasks, there’s no single best forecasting model. The real skill is knowing the strengths and limits of each approach and matching them to the problem.
Leave a Reply