In short: Model families differ by the data shapes they were designed for, so the pedigree behind a family is a better guide to fit than the length of a vendor's model list. Statistical families such as ARIMA, ETS and Theta were built for regular series with enough history to estimate seasonality, gradient boosted trees expect explanatory features, and deep and foundation models expect many related series. Four questions answered from your own history, covering series length, intermittency, the drivers you actually hold, and how many comparable series exist, remove most of a long list before anything is run. An ensemble tends to improve average accuracy while costing interpretability and runtime, which makes it a trade rather than a default.
Open the model picker in any modern planning platform and you get a wall of names. SARIMA, ETS, Theta, TBATS. LightGBM, XGBoost, CatBoost, Random Forest, ElasticNet. LSTM, DeepAR, N-BEATS, N-HiTS, TFT. Chronos, TimeGPT, Moirai, TimesFM, Lag-Llama, PatchTST, iTransformer. A Bayesian structural time series somewhere near the bottom, and a handful of research adapters behind a toggle that says experimental.
Our own catalogue has thirty-three entries across eight families. I want to be straight about what that number means, because vendors quote it as if breadth were the product. It isn't. Most of those models will never touch most of your catalogue, and a planner who is handed all thirty-three and told to choose has been given a problem rather than a tool.
What actually matters is the pedigree behind each family, and the fact that the pedigree tells you which data shapes each one was built for. Once you know that, the choice collapses to four questions you can answer from your own history.
Where each family came from
The oldest branch is statistical decomposition, and it is older than most people assume. Exponential smoothing dates to the late 1950s, in work by Brown, Holt and Winters, and the idea is unglamorous in a way that has aged extremely well: a weighted average where recent observations count more, extended to handle a trend and a season. Box and Jenkins formalised the autoregressive integrated moving average family in 1970, giving the field a systematic way to identify, estimate and check a model rather than guess at one. SARIMA is that machinery with a seasonal term bolted on.
Theta arrived in 2000, from Assimakopoulos and Nikolopoulos, and it won the M3 competition, which was the largest empirical forecasting comparison of its era. The method is almost embarrassingly simple. Decompose the series into two lines with different curvature, extrapolate each, average them. It has been beating far more elaborate approaches for a quarter of a century and it still shows up in the top ranks of comparisons it has no business being in.
The tabular machine learning branch grew out of a different tradition entirely. Gradient boosting comes from Friedman's work around 1999 and 2001, and the practical implementations arrived much later: XGBoost in 2016, LightGBM in 2017, CatBoost around the same time. These work outside the time series frame altogether. They treat a forecast as a supervised regression where the features happen to include lags, calendar variables, price, promotion flags and whatever else you can join in. That framing is exactly why they took over retail forecasting. Promotion response, cannibalisation, cross-price effects and holiday interaction are all just features, and a tree ensemble is very good at finding interactions between features without being told they exist.
The evidence here is independent of vendor claims. The M5 competition, run on Walmart data in 2020, was won by gradient boosted trees, and the same family dominated the upper reaches of the leaderboard. That result is the single most useful data point in this entire field for anyone planning consumer goods demand, because it was run on real hierarchical retail data at the item-store-day grain with promotions and price attached.
The deep learning branch is younger and its history is more mixed than the marketing suggests. Long short-term memory networks come from Hochreiter and Schmidhuber in 1997, though they were not applied seriously to demand until much later. DeepAR, from Amazon, first posted in 2017 and published in the International Journal of Forecasting in 2020, was the first widely used deep model that produced a probability distribution rather than a point, which mattered more than its accuracy did. N-BEATS in 2019 and N-HiTS in 2022 went the other way, using pure feedforward blocks with a basis expansion and no recurrence at all, and proved that a deep architecture could beat statistical baselines on the M4 dataset without any hand-engineered features. The Temporal Fusion Transformer from Lim and colleagues in 2021 combined attention with a variable selection network and quantile outputs, which is why it tends to be the deep model people actually keep when they run a bake-off.
The newest branch is foundation models. Chronos from Amazon in 2024, TimeGPT from Nixtla, Moirai from Salesforce, TimesFM from Google, Lag-Llama out of academia. These are pre-trained on enormous corpora of time series and can forecast a series they have never seen without being trained on it. That capability is real and it is genuinely new. The claims made about it are frequently ahead of the evidence, and independent evaluations have been mixed, with foundation models winning convincingly on some datasets and losing to a well-tuned gradient boosting model on others.
What the pedigree tells you about fit
Read that history again with an operational eye and a pattern falls out. Each family was built to exploit a different kind of signal, which means each one fails in a different way when that signal is absent.
Statistical models exploit the structure inside the series itself. They want a reasonable amount of history, a stable pattern, and not much else going on. Give SARIMA a mature product with a clean weekly rhythm and it will produce a defensible forecast in milliseconds that a planner can explain to a finance director without a slide. Give it a product whose demand is driven by a promotion calendar and it has no mechanism to see the promotion at all, so it will smooth straight through the spike and then treat the aftermath as a decline.
Tabular models exploit relationships between the series and everything around it. They want features. Price, promotion type, display, competitor activity, weather, calendar effects, item attributes. Given those, they are the strongest general choice in consumer goods and retail, and the competition record supports that. Take the features away and they degrade toward a mediocre autoregression, because you have removed the thing they were built to use.
Deep models exploit patterns across many related series at once. Their advantage shows up when you have hundreds or thousands of items that share behaviour, so that the model can learn a launch curve or a seasonal shape from the population and apply it to a series with thin history of its own. On a single series with two years of data they are usually a slow way to reach the same place a simpler method already got to.
Foundation models exploit what they learned somewhere else. That makes them genuinely useful in exactly one situation that has always been painful, which is a product with no history at all. A cold start is where a zero-shot forecast has no competition, because every other family needs data you do not have.
The four questions
Here is how the choice actually resolves, in the order the questions should be asked.
How intermittent is the series? Run the demand interval and variability classifier before anything else. If the item sells in fewer than about three quarters of periods and the demand sizes jump around, none of the above applies and you should be using Croston, the Syntetos-Boylan correction, or Teunter-Syntetos-Babai, which is the one that notices a product quietly dying. Fitting a transformer to a spare part that sells four times a year is a category error, and the fact that it will produce a number is not evidence that it worked.
Do you have real drivers, or only history? If price, promotion and calendar data are joined and trustworthy, a gradient boosting model is the default and the burden of proof sits with anything that wants to replace it. If all you have is the quantity column, the honest ranking flips and a well-fitted statistical model becomes hard to beat. This one question resolves a surprising share of model selection arguments, and it is answerable in an afternoon by looking at what is actually in the warehouse rather than what the integration document promised.
How much history, and how many related series? Under about two years, deep models have little to work with and pooling across items is the only thing that will save them. Above that, with a large catalogue of items behaving similarly, pooled deep models start to earn their cost. Note the second half of that condition. Long history on one lonely series is not the same thing as a large panel, and it does not unlock the same methods.
Is this item worth the compute? Most catalogues are steeply concentrated. A small share of items carries most of the volume and almost all of the margin exposure. Running an attention-based model with a variable selection network across a hundred thousand slow movers costs real money and buys almost nothing, because those items were never going to be forecastable at that grain in the first place. Segment first, then spend the expensive models where a point of accuracy converts into cash.
The ensemble question, and why it is not a free win
Every catalogue has a weighted ensemble at the bottom of the list, and combining forecasts is one of the most reliably useful findings in the whole literature. Forecast combination has been known to beat its own components since Bates and Granger in 1969, and the M4 competition was won by a hybrid that blended statistical and neural components.
The caveat is that an ensemble inherits every weakness its members share. If all six models in your blend are trained on the same censored sales history, the ensemble will be confidently wrong in exactly the same direction as each of them, and the fact that they agree will make it look more trustworthy rather than less. Combination protects you against model risk. It does nothing about data risk, and data risk is the larger of the two in most planning estates I have seen.
There is also a governance cost. A blend of six models is harder to explain in an S&OP review than any single one of them, and if the room cannot follow how a number was produced it will discount the number. That is a real constraint and it is worth pricing rather than dismissing as a communication problem.
What I would not claim
A model catalogue is only as good as the process that promotes things into it, and that process is where most of the value sits. Rolling origin evaluation rather than a single holdout, so a lucky fold cannot crown a model. A champion that a challenger has to beat by a margin that survives a sign test, so noise does not cause a promotion. Drift monitoring afterwards, because a model that won in March will quietly stop winning and nothing will announce it. Those mechanics are less interesting than the model names and they matter considerably more.
I would also not promise you a number. Published benchmark accuracy for any of these families is a property of the dataset it was measured on, and yours is different. The only accuracy figure worth anything is one measured on your own history against the method you are running today, which is a thing you can arrange before you buy anything.
Keeping thirty-three adapters rather than three is justified by the catalogue itself, and how many of them get used is a separate question. It lets you settle the argument empirically for each segment rather than by seniority, and the segments genuinely do disagree. Our promotional items and our long-tail spares want different families, and any single-model platform is quietly making that choice for you and calling it a philosophy.
Start with the four questions. They will eliminate most of the list before you have run anything.