Suggested visuals: share your screen with the poster or flip through Figures 1–4 from the report. Timing marks are approximate — practice once and adjust. Personalize the opening line before recording.
[0:00 – 0:35] Opening and the question
Hi, I’m David, and for my final project I looked at a question that sounds almost too simple: if you have thirty-four years of monthly data, how much does a real statistical model actually buy you over forecasting methods you could do with a pencil?
The dataset is U.S. retail sales at beer, wine, and liquor stores — FRED series MRTSSM4453USN, from the Census Bureau. Monthly, not seasonally adjusted, January 1992 through April 2026, 412 observations. I picked the non-adjusted version on purpose, because one of my four models is a seasonal naïve benchmark, and that model is pointless on data where the seasonality has already been removed.
[0:35 – 1:20] The data — show Figure 1
Two things jump out of the raw plot. First, the series has quadrupled since 1992 — there’s a strong persistent trend. Second, there’s a big December spike every year, and the spike gets bigger as the series grows. The STL decomposition confirms both: trend plus seasonality whose amplitude scales with the level. That second detail matters — it’s a hint that a multiplicative model should win, and it’s a hint the data gave us before we fit anything.
One more thing about the setup. I used a chronological 80/20 split, which puts the boundary at mid-2019. That means the entire COVID period lands in the test set. The models are trained on a world that ends in 2019 and judged on a world where everyone suddenly bought a lot more alcohol at retail in 2020. I didn’t plan that, but it turned out to be the most interesting thing about the project.
[1:20 – 2:20] The four models
The models, per the assignment: an ETS exponential smoothing model with automated selection — I fit the full grid of error, trend, and seasonal combinations and picked the winner by AICc; a naïve model that just repeats the last training value; a seasonal naïve model that repeats the last training year month by month; and an ensemble that averages the three.
The automated selection chose ETS(M,A,M) — multiplicative error, additive trend, multiplicative seasonality. Which is exactly what the decomposition predicted. I’ll admit I find it satisfying when the information criterion and the eyeball test agree.
All four models make one fixed-origin forecast 83 months ahead. No refitting along the way. That’s the harder version of the exercise, and I think the more honest one.
[2:20 – 3:30] Results — show Figures 2 through 4
The ranking is not close. ETS comes in with an RMSE of about 602 million dollars and a MAPE of 9.6 percent. The naïve model is at 1,208, seasonal naïve at 1,244 — both roughly double the error — and the ensemble lands in between at 974. ETS was also the only model with a positive out-of-sample R-squared. The naïve models were both below negative 1.6, meaning a flat line at the test-period average would have beaten them.
The error plot tells you why. Every model under-forecasts almost everywhere — the mean errors are all positive — because none of them saw COVID coming. The difference is what happens next. ETS extrapolates the trend, so it eventually catches up to the new higher level. The naïve models are frozen at 2019 and just fall further behind every year.
[3:30 – 4:25] The part that surprised me
Here’s what I keep coming back to: the ensemble result. The forecasting literature — Bates and Granger, the M4 competition — leans hard toward “combining forecasts wins.” And technically the average beat two of its three components. But it lost to ETS by a wide margin, because averaging only helps when the component errors partially cancel. Here, two of the three components shared the same structural bias, and the average inherited two-thirds of it. Combining works when the pieces know different things. These pieces mostly shared the same ignorance.
[4:25 – 5:00] Limitations and close
Honest caveats: this is one train-test split, not a rolling evaluation, so there are no error bars on any of this. The test window contains a once-in-a-century shock, so I would not quote 9.6 percent MAPE as a normal-year expectation. And the series is nominal dollars, so some of that trend is just inflation.
If I kept going, I’d try rolling-origin cross-validation, a weighted ensemble, and a dynamic regression with an explicit pandemic intervention term — my guess is that last one wins. Thanks for watching.
Recording notes: aim for 140–150 words per minute; the script runs ~4:50 at that pace, leaving buffer. Embed the finished video in the discussion board post per the assignment instructions.