In brief: Time series methods forecast future demand by examining values of past demand measured at the same intervals. The success of the method depends on correctly recognizing the structure of the series: demand may fluctuate around a certain level (stationary), increase or decrease over time (trending), or show fluctuations that repeat in certain periods (seasonal). The appropriate method differs for each structure; choosing the wrong method leads to systematic error in the forecast. This article covers the basic concepts of time series analysis, distinguishing stationary, trending, and seasonal series, the main methods, and points to watch in practice; it also examines the role of technology in collecting and analyzing data and interpreting errors. The general framework of forecasting methods can be reached on the What Is Demand Planning? page. In this article:What Is a Time Series and Why Is It Used in Forecasting?What Are the Components of a Time Series?How Are Stationary, Trending, and Seasonal Series Distinguished?What Does the Concept of Stationarity Mean?What Are the Main Time Series Methods?What Should Be Considered in Time Series Forecasting?How Does the Time Series Approach Change for Intermittent Demand?Technology and Digitalization: Collecting Data, Analysis, and Interpreting ErrorsImplementation: Checklist What Is a Time Series and Why Is It Used in Forecasting? A time series is the arrangement in time order of values of the same quantity measured at equal intervals (daily, weekly, monthly). In demand planning, this quantity is usually the sales or shipment quantity of a product or product family. The basic assumption of time series methods is that the behavior pattern of demand in the past will continue into the near future. The strength of this approach is that it needs only demand history and can be applied automatically to a large number of products. Its weakness is that it does not explicitly model the causes that affect demand (price, campaigns, economic conditions). If demand changes markedly with these factors, time series methods are supported with causal methods (such as regression) or with judgmental adjustments. For new products, since there is not enough history, time series methods cannot be used on their own. What Are the Components of a Time Series? A demand series can be thought of as a combination of several components: Level: The average demand quantity around which the series fluctuates.Trend: The movement of the series in the direction of increase or decrease over the long term. Market growth, the product life cycle, or pricing policy can cause a trend.Seasonality: Fluctuation that repeats in certain periods every year or every week. Beverage demand rising in summer months, food demand rising before Ramadan, or year-end rush are examples.Cyclical movement: Fluctuations that are usually tied to economic conditions, of non-fixed length, spreading over several years. The difference from seasonality is that its length cannot be known in advance.Random variation: Small fluctuations that cannot be explained or predicted. No method can eliminate this component; the purpose of the method is only to separate the other components correctly. How the components combine is also important. In the additive model, the size of the seasonal fluctuation is constant (200 units above average every July). In the multiplicative model, the fluctuation grows in proportion to the series level (20% above average every July). For products whose demand grows over time, the multiplicative model is more often appropriate. Choosing the wrong model leads to incorrect calculation of the seasonal effect as demand grows. How Are Stationary, Trending, and Seasonal Series Distinguished? A stationary series fluctuates randomly around a certain average; it has no marked trend or seasonal pattern. Consumer products in the maturity phase with stable demand can fall into this group. In the chart of the series, values stay within a horizontal band. A trending series continually rises or falls over time. A product in its growth phase or a product whose market is shrinking can be given as examples. In this series, simple average-based methods lag behind the trend and continually produce low or high forecasts. A seasonal series contains high and low periods that repeat at certain intervals. The fluctuation may depend on the months of the year, days of the week, or holiday periods. Methods that do not take seasonality into account produce error in the same direction in the same periods every year. In practice, most series carry trend and seasonality together. For this reason, the first step is to plot the series and overlay it by year; fluctuations of similar shape across years point to seasonality, and a persistent shift in the average points to trend. Method selection is made according to this structure: moving average and simple exponential smoothing for stationary series, Holt's method (exponential smoothing with trend) for trending series, and the Holt-Winters method or the decomposition approach for series in which trend and seasonality are both present. What Does the Concept of Stationarity Mean? Stationarity can be used in two different senses in time series analysis. In planning language, "stationary demand" refers to demand that fluctuates around a certain level and has no trend or seasonality. In statistical language, a stationary series is one whose mean, variability, and time-dependent relationship structure do not change over time. A series containing trend and seasonality is not stationary in this sense. The importance of statistical stationarity is that some models work reliably only on stationary series. Autoregressive Models (ARIMA) are the best-known example. A non-stationary series is made stationary by taking the difference of successive values (differencing) or by using seasonal differences, and the model is built on this transformed series. Exponential smoothing methods, in contrast, model trend and seasonality directly as components and so do not require the series to be made stationary in advance. The practical consequence of this distinction for the planner is that if the series is not stationary, a method that explicitly models trend and seasonality should be chosen instead of a simple average-based method, or the series should be made stationary with an appropriate transformation. What Are the Main Time Series Methods? Naive forecast: The forecast for the next period is the last actual value (in the seasonal naive method, it is the same period of the previous year). Because of its simplicity, it is useful as a comparison baseline for other methods. Moving average: The average of the last certain number of periods is used as the forecast. As the window lengthens, the forecast becomes smoother but reacts to changes later. It is appropriate only for stationary series. Simple exponential smoothing: The new forecast is a weighted combination of the previous forecast and the last actual: New forecast = Previous forecast + α × (Actual − Previous forecast). The smoothing coefficient α is between 0 and 1. For example, if the previous forecast is 100, the actual is 110, and α = 0.3, the new forecast is 100 + 0.3 × 10 = 103 (figures are for illustration purposes). A large α value reacts quickly to recent data but also follows random fluctuation; a small α value produces a more stable but slower forecast. Holt's method (exponential smoothing with trend): A method that updates the trend component in addition to the level. It is suitable for trending series. Since the assumption that the trend continues forever over long horizons can lead to high forecasts, the variant in which the trend is damped over time (damped trend) is frequently preferred. Holt-Winters method: A method that updates the level, trend, and seasonal components together. The seasonal component can be defined as additive or multiplicative. It is a common starting method for series containing trend and seasonality. Decomposition: The series is first separated into trend, seasonality, and remainder components. Seasonal indices are calculated, the series is deseasonalized and the trend is forecast, and the forecast is multiplied back by the seasonal indices. Because each component can be seen separately, it makes explanation easy. Autoregressive Models (ARIMA): Methods that model the relationships between the past values of the series and its past errors. It works on a series that has been made stationary and can capture complex autocorrelation structures. Model building requires expertise; automatic model selection reduces this burden. What Should Be Considered in Time Series Forecasting? Length and Quality of Historical Data To forecast seasonality reliably, data for at least two, preferably three complete seasons (24-36 months for monthly data) is needed. In shorter series, seasonal indices can mistake random fluctuation for seasonality. If the amount of data is insufficient, seasonality is used by deriving it from similar product groups or from a higher level. Cleansing and Flagging Events Time series methods learn the past as it is. Sales that stayed low because of stockouts, campaign peaks, one-off large orders, and system errors distort the structure of the series. These periods should be flagged as events in the series, corrected, or excluded from the model. Otherwise, campaign peaks are learned as normal demand and stockout periods as low demand. Calendar Effects Because religious holidays and Ramadan shift about 11 days earlier every year, seasonal indices tied to fixed months assign this effect to the wrong months; these periods should be handled as a separate event variable. In addition, the difference in the number of working days between months also creates fluctuation in the monthly series; working with demand per working day reduces this effect. Structural Changes Winning or losing a large new customer, a pricing policy change, a regulatory change, or a product renewal can permanently shift the level of the series. Data from before such breaks may not represent the new situation; it may be necessary to increase the weight of post-break data or to restart the model. Overfitting and Validation A model with many parameters can fit past data very well, but this success does not mean it will forecast the future well. In method selection, the model's success on data it has not seen, not its fit to the past, should be the basis. For this, the last section of the data is set aside, the model is built with the earlier section and tested on the section set aside; this process is repeated by rolling forward (rolling-origin validation). Forecast Horizon and Level As the forecast horizon lengthens, accuracy falls; a method that works well one week ahead may not be suitable for six months ahead. The method should be evaluated according to the horizon at which the decision is made. Level also matters: a series that is noisy at the product-location level looks much smoother at the product family level. Top-down and bottom-up forecasting approaches are therefore evaluated together. Regular Monitoring of Error No method stays accurate forever. When the structure of demand changes, the selected model also becomes outdated. For this reason, forecast error should be measured regularly, and if bias is consistently in the same direction, the method or parameters should be reviewed. For calculating and interpreting error, see How to Measure and Interpret Forecast Error: Bias and Percentage Error Metrics. How Does the Time Series Approach Change for Intermittent Demand? For some products (spare parts, special-order items), demand is zero in many periods and then arrives in variable quantities. These series cannot be modeled well with the methods above: zeros pull the average down, and exponential smoothing continually lowers the forecast in periods with no demand and raises it when demand arrives. In such series, the demand pattern is classified first. The average interval between demands (ADI) and the squared coefficient of variation of demand size (CV²) are used; the cut-off values common in the literature are 1.32 for ADI and 0.49 for CV². According to these values the series is called smooth, erratic, intermittent, or lumpy. For intermittent and lumpy series, Croston's method forecasts by smoothing the demand size and the interval between demands separately. The Syntetos-Boylan approximation (SBA), which reduces the upward bias of Croston's method, is also widely used. In these series, accuracy is evaluated with absolute error or service-level-based metrics instead of percentage error. For the application to spare parts in fields such as defense and aerospace, see Spare Parts Planning in Defense and Aerospace: Repair Cycle, Criticality, and Obsolescence. Technology and Digitalization: Collecting Data, Analysis, and Interpreting Errors Time series analysis can be done on a spreadsheet for a small product group. The difficulty is carrying out the same analysis consistently every period across thousands of product-location series. The contribution of technology can be considered in three stages: collecting and preparing data, analyzing series and selecting models, and calculating and interpreting errors. Collecting and Preparing Data Enterprise Resource Planning (ERP) systems are the source of order, shipment, inventory, and price transactions; the raw material of the time series is these records. Sales and inventory data received through Electronic Data Interchange (EDI) from distributors and retail chains, and consumption data from connected points of sale and smart shelves within the Internet of Things (IoT), produce series closer to true demand than shipments. This data is collected automatically from source systems and consolidated in a common data warehouse; a data warehouse is an analytical database in which data from different sources is stored in a single, consistent structure. Converting raw data into a time series is a preparation step in itself. Distinguishing missing periods from zero sales, unit conversions, merging product code changes, flagging stockout and campaign periods, and adding calendar variables can be done with automated data flows. When this step is automated, series are prepared with the same rules every period. Analysis of Series and Model Selection Statistical forecasting engines automatically classify trend and seasonality for each series, separate intermittent series, and choose the most suitable method among candidate methods. Automatic model selection is done by comparing candidates' success on unseen data (rolling-origin validation) or model information criteria. Thus, expert evaluation of thousands of series one by one is not needed; the expert works only on important and problematic series. Artificial Intelligence (AI) is the general name for methods that produce forecasts, classifications, and recommendations by learning patterns from data; Machine Learning (ML) is its sub-field that learns from data. Tree-based models and neural networks can learn many series together and benefit from the patterns of other series for products with short history; they can also easily add external variables such as price, weather, and calendar. Pre-trained general-purpose time series models have also come into use. All of these models should be subjected to the same validation discipline as classical methods: if a new model does not give measurably better results than the existing statistical method, it is not adopted into the process. Calculating and Interpreting Errors When forecast and actual are stored in the same structure, bias and percentage error metrics can be calculated automatically across product, product family, and channel breakdowns. Business Intelligence (BI) tools are the analytics layer that combines data from different sources and presents it to decision makers through dashboards and reports; error heat maps, forecast-versus-actual charts per series, and exception lists of items exceeding a threshold are presented in this layer. Error is also an early indicator that the model has become outdated; series whose bias becomes persistent are flagged for model re-evaluation. Generative Artificial Intelligence (GenAI) is a type of artificial intelligence that produces summaries and draft explanations in natural language based on large language models. It can turn error findings into a readable assessment; however, verifying the likely causes and making the decision remain the planner's responsibility. A time series method learns the past but does not know the cause; knowledge of the cause comes from the planner and the sales team Implementation: Checklist Plot the chart for each series; verify trend, seasonality, and intermittent structure visually.Flag and clean stockout, campaign, and one-off event periods in the series.Use data from at least two, preferably three complete seasons for seasonality.Model Ramadan, holiday, and working day effects with separate variables.In method selection, base the decision on success on unseen data, not fit to the past.Monitor forecast error and bias regularly; re-evaluate the model for series that consistently deviate in the same direction.Consolidate ERP, point-of-sale, and shipment data in a common data warehouse; automate the series preparation steps.Before putting machine learning models into operation, subject them to the same validation as the existing statistical method. Related Articles How to Measure and Interpret Forecast Error: Bias and Percentage Error MetricsHow to Forecast Promotion Impact: Campaign Types and Uplift Calculation5 Common Mistakes in Safety Stock Calculation and the Correct ApproachSpare Parts Planning in Defense and Aerospace: Repair Cycle, Criticality, and Obsolescence Frequently Asked Questions When are time series methods not sufficient? Time series methods alone are insufficient if demand changes markedly with price, campaigns, or external factors, if there is not enough history for new products, or if the structure of demand has changed permanently. In these cases they should be supported with causal methods and judgmental knowledge. How many periods of historical data are needed? For stationary series, a few dozen observations are often sufficient. For forecasting seasonality, at least two, preferably three complete seasons are needed; for monthly data this corresponds to 24-36 months. Why is stationarity important in planning? Whether the series is stationary determines method selection. When a simple method designed for stationary series is applied to a series containing trend or seasonality, the forecast deviates consistently in the same direction. How is the exponential smoothing coefficient chosen? The coefficient is determined on past data so as to minimize error in periods the model has not seen. Small values are preferred for products with stable demand and large values for products that need to react quickly to change. Modern forecasting engines make this selection automatically. Will artificial intelligence replace classical time series methods? No. Artificial intelligence models may give better results on some series, but they are mostly used together with classical methods and compared with the same validation. A reliable decision rests on showing that the model adds measurable value. References Hyndman, R. J., & Athanasopoulos, G. (2021). Forecasting: Principles and Practice (3rd ed.). OTexts.Makridakis, S., Wheelwright, S. C., & Hyndman, R. J. (1998). Forecasting: Methods and Applications (3rd ed.). Wiley.Box, G. E. P., Jenkins, G. M., Reinsel, G. C., & Ljung, G. M. (2015). Time Series Analysis: Forecasting and Control (5th ed.). Wiley.Winters, P. R. (1960). Forecasting sales by exponentially weighted moving averages. Management Science, 6(3), 324-342.Gardner, E. S., & McKenzie, E. (1985). Forecasting trends in time series. Management Science, 31(10), 1237-1246.Croston, J. D. (1972). Forecasting and stock control for intermittent demands. Operational Research Quarterly, 23(3), 289-303.Syntetos, A. A., Boylan, J. E., & Croston, J. D. (2005). On the categorization of demand patterns. Journal of the Operational Research Society, 56(5), 495-503. Related Optimum Planlama Trainings finance_mode Demand Planning and Sales Forecasting Learn forecasting processes, methods, and analyses through hands-on applications. Detailed Information package_2 Inventory Management and Optimization Classify inventory and optimize inventory quantities and costs. Detailed Information Companies That Trust UsAkkim, 2023Alarko Holding, 2011Anadolu Efes, 2021Barilla Food, 2015BVN (Bahcivan), 2020Cilek Furniture, 2017Civil Retail, 2019Eczacibasi Esan, 2021EssilorLuxottica, 2022Global Bilgi, 2021Global Pharmaceutical Company, 2023Haribo, 2017Hascelik, 2015, 2022Ibavalresa, 2025Istikbal Furniture, 2012Jantsa, 2019Kaleseramik, 2015Kastamonu Entegre, 2017LCWaikiki, 2010, 2012Makbul Nuts, 2021, 2022Migros, 2021Nivea Türkiye, 2019Opet Fuchs, 2014Panasonic Electric Works, 2022Red Bull Türkiye, 2023, 2025Rengarenk Furniture, 2017Sarsilmaz Firearms, 2023, 2024Sika Türkiye, 2018Sisecam Holding, 2013–2016SKT Spare Parts, 2019, 2023Turk Telekom, 2013–2018Zorluteks, 2023Comments “We had a productive training for two days. As a company, we have seized opportunities to see our shortcomings and make improvements. Thank you very much.” Yiğit Doruk Turguter, Senior Engineer in Production and Demand Planning, Petrol Ofisi (Demand Planning, 2024) “It was a training that was useful for our sales and planning group. Thank you.” Murat Işık, Planning & ERP Manager, Sarsilmaz Firearms (Demand Planning, 2023) “It is a training that every company operating in a geography like Turkey, where uncertainty is increasing day by day, should receive. I think that this training will provide a great advantage in our business processes, as it is very critical to make the right decisions in an aggressive competitive environment and to make the right predictions in parallel.” Hale Topaloğlu, Senior Product Manager, Global Pharmaceutical Company (Demand Planning, 2023) To view all testimonials, click here. Request a quote for training or consulting tailored to your organization. Request a Quote