A time series is a sequence of values ordered in time, where the past illuminates the future. Forecasting them is the mother discipline of demand planning. It is also one of the rare fields where open scientific competitions, the M competitions, have established on real data what truly works, upending many certainties along the way.
Why it matters
Almost every demand forecast is, at bottom, a time-series forecast: you observe a reference's sales history and try to anticipate what comes next. What sets a time series apart from a mere scatter of points is order: each value arrives after the previous one and often depends on it. This dependence on time, autocorrelation, is both the central difficulty and the main resource: it is because the recent past informs the near future that forecasting is possible at all.
Understanding time series means understanding the anatomy of this dependence. A series classically decomposes into an underlying trend, a repeating seasonal component, possible longer cycles, and an unpredictable residual noise. Each forecasting method is a particular way of capturing some of these components and extrapolating. A forecast's quality depends less on the tool's sophistication than on the fit between the series' real structure and what the model can capture.
This field has a rare feature in management: an objective referee. Since 1982, the M competitions organized by Spyros Makridakis have pitted, on tens of thousands of real series, every existing method against one another, from the simplest to the most advanced. Their published, reproducible results settle debates that intuition and commercial interests would leave forever open. They are the best available compass for knowing which method truly deserves its reputation.
The M competitions have delivered counter-intuitive verdicts that shaped the discipline. For decades, their most famous and unsettling finding was that simple methods are often as accurate as complex ones, and that combining several methods almost always beats the best single model. Sophistication for its own sake does not pay; robustness and combination do.
Then came the shift. Until M4 (2018), pure machine learning disappointed, beaten by statistical approaches and their combinations; only a statistical-neural hybrid won. At M5 (2020), on real Walmart data, the shift was clear: gradient boosting (LightGBM) and global learning across thousands of series at once finally dominated, beating the best statistical benchmarks by more than 20%. The expert lesson is not 'ML wins', but subtler: ML wins when there are many related series to learn together, not on an isolated series. The context makes the method.
The mechanism
Forecasting a time series always follows the same logic: understand the past's structure (trend, season, cycle, noise), choose a model that captures that structure, then extrapolate it into the future while quantifying uncertainty.
The first step is always diagnostic: does the series have a trend? a seasonality, and of what period? is its variability stable or growing? is it continuous or intermittent? This diagnosis points to the right family of methods. An error here, like ignoring a marked season, dooms the forecast whatever compute power is thrown at it afterward.
Then comes the model choice, and this is where the M competitions light the way. The chart shows the three great families, in the order they rose to prominence. Local statistical methods (exponential smoothing, ARIMA) fit one model per series: robust, frugal, legible, they remain excellent defaults and formidable benchmarks. Global machine learning methods (gradient boosting, neural networks) learn across thousands of series at once and capture patterns no isolated series reveals. Hybrid and combined approaches, finally, fuse both worlds, and they are the ones topping the rankings.
At the summit of the state of the art today are specialized architectures: N-BEATS and DeepAR (deep neural networks designed for series), meta-learning methods like FFORMA that learn to intelligently weight an ensemble of models by each series' features, and lately foundation models pre-trained on vast corpora of series, able to forecast a never-seen series without retraining. These advanced methods do not replace diagnosis: they reward it, exploiting a correctly identified structure better.
Figure 1. The three families of methods, in the order the M competitions saw them rise: from local statistical to global ML, to the hybrid approaches that dominate the state of the art. Illustrative schematic.
The traps
The M competitions also revealed the traps that ensnare those who neglect method for tooling.
The biggest lesson of the M competitions is that a poorly chosen sophisticated model loses to a well-fitted simple one. Deploying a deep neural network on a handful of short series, where exponential smoothing would do, costs much and forecasts worse. Complexity is justified only by the data's structure, never by itself.
Betting on a single model, even the best on average, leaves accuracy on the table. The M competitions show, repeatedly, that combining several methods reduces error and above all stabilizes it. Refusing combination for love of a single model's elegance is a luxury accuracy does not forgive.
Evaluating a model by mixing past and future (classic cross-validation) gives illusory performance, because you leak future information into the past. A time series is validated by moving forward in time (rolling backtesting), never by drawing points at random. Ignoring this rule means lying to yourself about true quality.
The methods
Choosing and deploying a time-series method follows a progression the M competitions validated on real data.
Characterize trend, seasonality, variability, intermittency, and the amount of available data. This diagnosis, more than anything else, determines the relevant family of methods. It always precedes the choice of algorithm.
First establish the performance of simple, robust methods (seasonal naïve, exponential smoothing, ARIMA). No advanced method deserves deployment unless it clearly beats these references. It is the basic discipline of the M competitions.
If you have many related series (a large portfolio of references), consider global learning (gradient boosting, deep networks like N-BEATS or DeepAR) that learns from one series to another. That is where, and mainly where, ML gains the edge.
Fusing several models (averaging, meta-learning like FFORMA) almost always improves and stabilizes the forecast. Validate the whole by rolling backtesting that respects time order, measuring error with a cross-series comparable metric like MASE.
Neighboring concepts
Time series encompass smoothing, season and trend, and lead into the probabilistic.
From knowledge to action
Between simple benchmarks, global ML and combination, the right choice depends on your data. Our Planning, Forecasting & S&OP file aligns your methods with what science has validated.