Lesson 1 - Simple Exponential Smoothing

Welcome to Simple Exponential Smoothing#

Modules 5 and 6 built one whole forecasting family around autocorrelation: AR, MA, ARMA, ARIMA, and SARIMA. This module builds a second, different family. Exponential smoothing does not model autocorrelation directly. Instead, it forecasts with a weighted average of past values, where recent values get more weight and older ones get exponentially less. Simple exponential smoothing (SES) is the simplest member: it assumes the series has a level that can shift, but no trend and no season.

By the end of this lesson, you will be able to:

• Write the SES update formula and explain what the smoothing parameter alpha controls
• Compute a few steps of SES by hand
• Fit SES with statsmodels and read the fitted alpha
• Explain why SES fit to a trending series reduces to the naive forecast

Let’s build the simplest smoother first.

The SES Formula#

SES keeps a single running estimate of the series’ level, called $l_t$. Each new observation nudges that level a little, and the forecast for every future point is just the most recent level:

The parameter $\alpha$ (alpha), between 0 and 1, controls the trade-off. A high $\alpha$ means the level moves quickly toward the newest observation, trusting recent data heavily. A low $\alpha$ means the level barely moves, trusting the long history built up so far. Because each update mixes in the previous level, and that previous level already mixed in the one before it, the influence of any single past observation shrinks by a factor of $(1 - \alpha)$ every step after it arrives. That is the “exponential” in exponential smoothing: the weights on past observations decay exponentially, not the forecast itself.

A Manual Example#

Six observations, and a fixed $\alpha = 0.3$, computed by hand:

obs = [50, 54, 49, 53, 51, 55]
alpha = 0.3

level = obs[0]
levels = [level]
for x in obs[1:]:
    level = alpha * x + (1 - alpha) * level
    levels.append(round(level, 3))

print(levels)
# [50, 51.2, 50.54, 51.278, 51.195, 52.336]

The level starts at the first observation (50), then drifts toward each new value by 30% of the gap each step. After 54 arrives, the level moves from 50 to $0.3 \times 54 + 0.7 \times 50 = 51.2$. After 49 arrives, it moves back down to 50.54. It never jumps all the way to the newest value. It always takes a fraction $\alpha$ of the step. Forecasting from here for any number of steps ahead just repeats the last level, 52.336, forever. A flat line is the only shape SES can produce, because it has no mechanism to represent a trend or a season.

What Alpha Means, on Two Toy Series#

Rather than guessing $\alpha$, you let statsmodels fit it by minimizing forecast error on the training data. Two small toy series show the two extremes:

import numpy as np
import pandas as pd
from statsmodels.tsa.holtwinters import SimpleExpSmoothing

rng = np.random.default_rng(5)
pure_noise = pd.Series(50 + rng.normal(0, 3, 60))

ses_noise = SimpleExpSmoothing(pure_noise, initialization_method="estimated").fit()
print(round(ses_noise.params["smoothing_level"], 4))   # 0.0
print(np.round(ses_noise.forecast(5).values, 2))        # [49.36 49.36 49.36 49.36 49.36]

rng2 = np.random.default_rng(9)
steps = rng2.normal(0, 5, 60)
drifting = pd.Series(50 + np.cumsum(steps) * 0.3 + rng2.normal(0, 1, 60))

ses_drift = SimpleExpSmoothing(drifting, initialization_method="estimated").fit()
print(round(ses_drift.params["smoothing_level"], 4))    # 0.5513

The first series is pure noise scattered around a constant level, with nothing real to track. The fitted $\alpha$ is 0.0: the optimizer finds that the best forecast is just the historical average, ignoring every individual fluctuation, because reacting to noise only adds error. The second series genuinely wanders, with each step building on the last. The fitted $\alpha$ is 0.5513, roughly splitting the difference between trusting history and trusting the newest point, because here the newest point really does carry information the average would miss. Reading a fitted $\alpha$ is reading how much real, trackable movement the optimizer found in the series.

Applying SES to Cyclepath#

Now fit SES on the real, trending, seasonal Cyclepath series and see what happens:

def cyclepath():
    idx = pd.date_range("2016-01-01", periods=96, freq="MS")
    t = np.arange(96)
    rng = np.random.default_rng(42)
    trend = 9000 + 90 * t
    seasonal = 3200 * np.sin(2 * np.pi * (t - 3) / 12)
    noise = rng.normal(0, 350, 96)
    return pd.Series(np.round(trend + seasonal + noise).astype(int), index=idx, name="trips")

y = cyclepath()
train, test = y.iloc[:-12], y.iloc[-12:]

ses = SimpleExpSmoothing(train, initialization_method="estimated").fit()
print(round(ses.params["smoothing_level"], 4))   # 1.0
print(round(ses.forecast(12).iloc[0], 1))          # 13565.0

The fitted $\alpha$ is exactly 1.0, the opposite extreme from the pure-noise toy series. At $\alpha = 1$, the SES update collapses to $l_t = y_t$: no smoothing at all, just the latest observed value carried forward as the forecast for every future month. Compare that to the naive baseline from Module 1, which forecasts every future month as the last observed value:

def mape(a, f): return np.mean(np.abs((a - f) / a)) * 100

naive = pd.Series(train.iloc[-1], index=test.index)
ses_fc = ses.forecast(12)

print(np.allclose(ses_fc.values, naive.values))   # True
print(round(mape(test, ses_fc), 2))                 # 18.99

The two forecasts are numerically identical, and SES’s test MAPE (18.99%) matches Module 1’s naive baseline (19.0%) to within rounding. This is not a coincidence, and it is not a bug. Cyclepath has a real trend, and SES has no term for a trend. Given a level-only model and a series that keeps climbing, the optimizer’s best move is to abandon smoothing entirely and just track the latest point, which is exactly what the naive forecast does. SES did not fail by producing something worse than naive. It failed by discovering that naive is the best a level-only model can do here.

Practice Exercises#

Exercise 1: Predict the next level by hand#

Continuing the manual example ($\alpha = 0.3$, last level 52.336), a seventh observation of 58 arrives. What is the new level?

Exercise 2: Interpret alpha near 0.9#

You fit SES to a series and get $\alpha = 0.92$. What does that tell you about the series, and what would the forecast look like?

Exercise 3: Why not just always use naive, then?#

If SES on a trending series just becomes the naive forecast, why bother fitting SES at all instead of using the one-line naive forecast directly?

Summary#

Simple exponential smoothing forecasts a single smoothed level, updated by $l_t = \alpha y_t + (1-\alpha) l_{t-1}$, with every future forecast equal to the most recent level, a flat line. On a pure-noise toy series, the fitted $\alpha$ went to 0.0, trusting the historical average over any single fluctuation. On a genuinely drifting toy series, it landed at 0.5513, splitting the difference. On Cyclepath, a series with a real trend and no trend term available, the fitted $\alpha$ went to the opposite extreme, 1.0, which collapses the update to $l_t = y_t$ and makes the SES forecast numerically identical to the naive baseline from Module 1: same values, 18.99% MAPE versus the naive baseline’s 19.0%.

Key Concepts#

• SES level update — $l_t = \alpha y_t + (1-\alpha) l_{t-1}$; the forecast is the most recent level, repeated flat forever.
• Alpha as a dial — near 0 trusts history (the average), near 1 trusts only the newest observation.
• Exponentially decaying weights — each past observation’s influence shrinks by $(1-\alpha)$ every step after it arrives.
• Alpha = 1 is naive — on a series with a real trend and no trend term, SES optimizes to the naive forecast.

Why This Matters#

A fitted alpha is not just a number to report, it is a diagnostic. An alpha near 0 says the series is close to noise around a constant. An alpha near 1 says the model found nothing worth smoothing and gave up on history entirely, which is exactly what happens when a real trend has no term to attach to. Recognizing that failure mode here, rather than being surprised by it later, is what makes the next two lessons feel like a direct response to a problem you have already seen, rather than new material appearing out of nowhere. Next, Holt’s method adds a trend term directly, and you will see it fail in an even more informative way on this same series.

Next Steps#

Continue Building Your Skills#

You have seen SES reduce to the naive forecast on a trending series, and you know why: alpha went to 1 because a level-only model has nothing else to offer a series that keeps climbing. Next, Holt’s method gives the model an explicit trend term. You might expect that to fix things immediately. Instead, on this seasonal series, it produces a forecast that is dramatically worse, for a reason that is just as informative as this lesson’s naive-equivalent result.