Types of forecasts

From CEOpedia
Types of forecasts
See also


Types of forecasts - is the classification of forecasts according to the criteria that matter to the people who make and use them: how far ahead they look, what they predict, which method produces them, what form the result takes, what purpose they serve and at which level of aggregation they are prepared. Hyndman and Athanasopoulos open their textbook with the plain observation that modern organisations require short-term, medium-term and long-term forecasts, depending on the specific application[1]. Horizon is only the first cut, though. The same sales figure for next March can be a point forecast or an interval, a judgment or a model output, a conditional "what if" or an unconditional projection.

Why bother sorting forecasts at all? Because each type answers a different question and fails in a different way. A planner who knows that she is holding a short-term, quantitative, point forecast of store-level demand also knows what it can't tell her: how wide the uncertainty is, or what happens if a competitor cuts prices. The general logic of prediction is covered in the article on forecasting. This one is about the map.

Classification by time horizon

The oldest and most common split. Boundaries differ between industries (a "long" horizon in fashion retail is a short one in power generation), but the usual reading in business is:

  • Short-term forecasts - days to a few months. They feed scheduling of personnel, production and transport, and day-to-day replenishment[1]. They're the most accurate of the three because the near future tends to resemble the recent past.
  • Medium-term forecasts - roughly a few months to two years. Used to decide resource requirements: buying raw materials, hiring staff, ordering machinery. Typical inputs to production planning and annual budgets.
  • Long-term forecasts - several years and more. These support strategic planning, where decisions must account for market opportunities, environmental factors and internal resources. Accuracy is low. Usefulness, oddly enough, can still be high, since the point is often to frame options rather than to hit a number.

Some authors add immediate (hours, minutes) forecasts for electricity load, call centres or high-frequency trading. Others merge medium and long into one "planning horizon". The labels matter less than the habit of matching the horizon to the decision lead time: there's no use forecasting demand two days ahead if a supplier needs six weeks to deliver.

Classification by subject

What is being forecast shapes the data available and the methods that make sense.

  • Economic forecasts - GDP, inflation, interest rates, unemployment, exchange rates. Produced by central banks, ministries, international organisations and private research houses, and consumed by firms as inputs to their own plans and to macro environment analysis.
  • Demand and sales forecasts - expected market demand or company sales by product, region, customer or channel. The workhorse type in management, sitting under sales planning, inventory policy and capacity decisions.
  • Technological forecasts - when a technology will mature, how fast it will diffuse, what performance it will reach. Methods lean towards expert panels, trend extrapolation of performance curves and technology scouting.
  • Financial forecasts - revenue, costs, cash flow, working capital, financing needs. They translate the other forecasts into money and are the backbone of financial planning.
  • Weather and climate forecasts - not a management product, but a major input for energy, agriculture, retail, insurance and logistics. Numerical weather prediction is also where probabilistic forecasting matured, through ensemble methods.
  • Workforce forecasts - future headcount, skills, turnover and retirements, often combined with analysis of labour market trends.

Other subject types exist (political, demographic, social). The list is open.

Classification by method

This criterion produces the richest typology. Armstrong's handbook organises forecasting methods into a tree whose first branch separates judgmental methods from statistical ones[2]. Most textbooks follow a similar three-way split.

Qualitative (judgmental) forecasts

Used when data are missing, irrelevant or about to be overtaken by a structural break: a new product, a new market, a regulatory change. Hyndman and Athanasopoulos stress that these methods are not purely guesswork, since structured approaches exist for getting good forecasts without historical data[3].

  • Delphi method - experts answer anonymous questionnaires in several rounds, see a summary of the group's answers and may revise their own. Rowe and Wright identify four defining features: anonymity, iteration, controlled feedback and statistical aggregation of the group response[4]. The technique was developed at the RAND Corporation in the 1950s by Olaf Helmer and Norman Dalkey.
  • Executive opinion (jury of executives) - senior managers produce an aggregate forecast, usually in a meeting. Fast. Prone to dominance of the loudest voice.
  • Sales force composite - salespeople or branch managers forecast their own territories and the numbers are summed. They know their customers best; they also have an incentive to under-forecast when targets are set from those same numbers[3].
  • Market research and customer intentions - surveys, panels and concept tests, both qualitative and quantitative. The catch is the gap between what people say they'll buy and what they actually buy.
  • Forecasting by analogy - borrowing the sales curve of a similar product launched earlier.

Judgmental forecasts usually come as subjective probabilities or single numbers, rarely as full distributions.

Quantitative time series forecasts

These use only the past values of the variable itself. The assumption is blunt: patterns observed so far will continue.

  • Naive forecast - next value equals the last observed value. Simple. Hard to beat for many economic and financial series, which is why it's the standard benchmark[1].
  • Moving average - the mean of the last k observations. Smooths noise, lags behind trends.
  • Exponential smoothing - a weighted average in which weights decline geometrically into the past. Proposed in the late 1950s by Brown, Holt and Winters, it grew into a family covering level, trend and seasonality[5].
  • ARIMA models - autoregressive integrated moving average models popularised by Box and Jenkins around 1970. They describe the autocorrelation structure of a series; the autoregressive model is the simplest member.
  • Decomposition and trend projection - separating trend, seasonal and irregular components, then projecting each.

Causal (econometric) forecasts

Here the forecast variable is explained by other variables: price, advertising, income, temperature, the number of working days. Tools range from simple linear regression to multiple regression and systems of simultaneous equations used in macroeconomic models. The strength is explanation. The weakness is that a causal forecast needs forecasts of its drivers, which brings us to the next criterion.

Classification by purpose and conditionality

Unconditional forecasts state what is expected to happen. Conditional forecasts state what will happen if certain assumptions hold. Every causal model run with assumed future values of its predictors produces a conditional forecast.

Hyndman and Athanasopoulos make a related distinction for regression models. Ex-ante forecasts use only information available at the moment of forecasting, so the predictors must themselves be forecast. Ex-post forecasts use later, actual values of the predictors (but not of the forecast variable) and serve to study how the model behaves. A third variant, scenario-based forecasting, plugs in values for the predictors chosen by the analyst: what happens to consumption if income grows 1% and unemployment stays flat?[6] This is the bridge to scenario analysis and sensitivity analysis.

A distinction found mostly in European textbooks separates passive forecasts, which assume the forecaster's organisation won't change the course of events, from active forecasts, which include planned actions (a price cut, a campaign, a new plant) as part of the picture. Close to this are warning forecasts, published precisely so that someone acts to prevent the predicted outcome, and their mirror image, self-fulfilling forecasts. The exchange rate example in Hyndman and Athanasopoulos fits here: if people believe a forecast that the rate will rise, they adjust prices at once and the forecast fulfils itself[1].

Classification by form

Same horizon, same method, three very different outputs:

  • Point forecast - a single number: "we'll sell 1,250 units in May". Usually the mean or median of the forecast distribution.
  • Interval forecast - a range with a stated probability: "between 1,100 and 1,400 units with 80% probability". The interval widens as the horizon lengthens.
  • Probabilistic (density) forecast - the whole probability distribution of future values, or probabilities attached to events ("30% chance of rain"). Gneiting and Katzfuss describe the goal as maximising the sharpness of the predictive distributions, subject to calibration[7]. Put simply: be as precise as you can, but your 80% intervals should contain the outcome about 80% of the time.

Point forecasts still dominate business practice. Thing is, inventory and risk decisions depend on the tails, not the centre. Safety stock is sized from the spread of demand, not its mean. The M4 Competition, which compared 61 methods on 100,000 time series, deliberately evaluated prediction intervals alongside point forecasts for this reason[8]. Simulation tools such as Monte Carlo simulation are a common way of turning a model into a full distribution.

Classification by level of aggregation

  • Macro forecasts concern the economy, an industry or a whole market. Micro forecasts concern one firm, product, store or customer. The two link up: a retailer's sales forecast often starts from a macro forecast of consumer spending.
  • Top-down forecasts are made for the total (all products, the whole country) and split into parts using historical proportions. Bottom-up forecasts are made for each item at the lowest level and summed. A middle-out approach starts at an intermediate level and goes both ways[9].

Bottom-up keeps detail but piles up noise from thousands of small series. Top-down is stable but blind to item-level shifts. Modern practice reconciles forecasts made at all levels so that they add up, the same tension that appears between top-down and bottom-up budgeting.

Summary table of types, methods and horizons

Forecast type Typical methods Typical horizon Typical users
Operational demand (SKU, store) Naive, moving average, exponential smoothing Days to weeks Inventory, logistics, store managers
Sales by product line or region Exponential smoothing with trend and seasonality, ARIMA, sales force composite Months to 1-2 years Sales, production planning, S&OP
Financial (revenue, cash flow) Regression on sales drivers, budgets, scenario forecasts Quarter to 3-5 years Finance, controlling, lenders
Economic (GDP, inflation, rates) Econometric models, ARIMA, expert surveys Quarters to several years Governments, central banks, strategy teams
New product / new market Delphi, executive opinion, customer intentions, analogy 1-5 years Marketing, product management
Technological Delphi, trend extrapolation, scenario analysis 5-20+ years R&D, strategic planning
Workforce Ratio and trend methods, regression on activity drivers, expert judgment 1-5 years HR, workforce planning
Weather / energy load Numerical models, ensembles, probabilistic forecasts Hours to 2 weeks Utilities, transport, retail

The boundaries are soft. A good forecasting system mixes types: a statistical baseline, a judgmental adjustment, an interval around both.

Example of comparing time series forecast types

Hypothetical data. A distributor records monthly sales of one product (units): 120, 132, 125, 140, 138, 150, 145, 160, 152, 148, 165, 170. Three methods are compared:

  • Naive: \( F_{t+1} = A_t \)
  • 3-period moving average: \( F_{t+1} = \frac{A_t + A_{t-1} + A_{t-2}}{3} \)
  • Simple exponential smoothing with α = 0.3, started with \( F_2 = A_1 \):

\[ F_{t+1} = \alpha A_t + (1-\alpha) F_t \]

Accuracy is measured over months 4-12 (the first month where all three methods have a forecast):

\[ MAE = \frac{1}{n}\sum |A_t - F_t| \qquad MAPE = \frac{100\%}{n}\sum \frac{|A_t - F_t|}{A_t} \]

Month Actual Naive MA(3) SES α=0.3
1 120 - - -
2 132 120 - 120.00
3 125 132 - 123.60
4 140 125 125.67 124.02
5 138 140 132.33 128.81
6 150 138 134.33 131.57
7 145 150 142.67 137.10
8 160 145 144.33 139.47
9 152 160 151.67 145.63
10 148 152 152.33 147.54
11 165 148 153.33 147.68
12 170 165 155.00 152.87
13 (forecast) ? 170 161.00 158.01

For month 4, for example, the moving average is (120 + 132 + 125) / 3 = 125.67, and the SES value is 0.3 × 125 + 0.7 × 123.60 = 124.02.

Method MAE (units) MAPE Mean error (bias)
Naive 9.22 6.02% +5.00
MA(3) 9.44 6.14% +8.48
SES α=0.3 12.59 8.19% +12.59

The naive forecast wins. Surprising? Not really. The series has a clear upward trend of roughly 4-5 units a month, and both averaging methods drag old, lower values into every forecast. Exponential smoothing with a small α reacts slowly; it under-forecast in every single month from 4 to 12, which is why its mean error equals its MAE. That's a systematic error, not noise. The fix is a method built for trend (Holt's linear method, or a regression on time), not a higher α picked by trial and error. Lesson: the "type" of a time series method has to match the pattern in the data, and a mean absolute deviation figure alone hides bias, so the sign of the errors should be checked too.

Advantages and limitations of the main families

Qualitative forecasts

Advantages:

  • Work without data, so they fit new products, new markets and long horizons.
  • Capture information no model sees: a competitor's plans, a coming regulation, a key customer's mood.
  • Structured versions such as Delphi reduce the influence of dominant individuals[4].

Limitations:

  • Optimism, anchoring and political bias. Sales targets and forecasts get mixed up.
  • Hard to reproduce or audit. Two panels can give two answers.
  • Slow and costly when repeated for thousands of items.

Time series forecasts

Advantages:

  • Cheap, fast and automatic, ideal for large numbers of items in short-term planning.
  • Simple methods are surprisingly competitive; that's why naive and exponential smoothing remain benchmarks[5].
  • Easy to monitor and update every period.

Limitations:

  • Blind to causes. They can't answer "what if we raise the price?"
  • Break down at structural changes: a pandemic, a new competitor, a lost key account.
  • Accuracy falls quickly with the horizon.

Causal forecasts

Advantages:

  • Explain why the forecast variable moves and allow conditional and scenario forecasts[6].
  • Support policy and pricing decisions, not just planning.

Limitations:

  • Need forecasts of the explanatory variables, which may be harder to predict than the target itself.
  • Correlation found in past data doesn't guarantee a stable relationship in the future.
  • More data, more expertise, more maintenance.

Point versus interval and probabilistic forecasts

Point forecasts are easy to communicate and feed straight into plans. But they hide uncertainty, and users tend to treat them as promises. Intervals and densities are more honest and more useful for risk and inventory decisions, but they're harder to explain and to evaluate. A further warning for the example above: MAPE is undefined when actual values are zero and explodes when they're close to zero, so intermittent demand needs scaled measures instead[10].

Related approaches

  • Combination of forecasts - averaging forecasts from different types often beats the best single one, which is one reason classification matters: you can only combine methods you've told apart.
  • Judgmental adjustment of statistical forecasts - common in practice, useful when based on real new information, harmful when done out of habit[3].
  • Scenario analysis and foresight - not forecasts in the strict sense, since they describe several possible futures without claiming which one is most likely.
  • Planning - forecasts describe what will probably happen; plans state what the organisation intends to do. The two should be kept apart, as different types of planning need different forecast inputs.

References

Footnotes

  1. ↑ Hyndman R.J., Athanasopoulos G., Forecasting: Principles and Practice, 3rd edition, OTexts, 2021, Chapter 1 and section 5.2
  2. ↑ Armstrong J.S. (ed.), Principles of Forecasting: A Handbook for Researchers and Practitioners, Springer, 2001
  3. ↑ Hyndman R.J., Athanasopoulos G., Forecasting: Principles and Practice, 3rd edition, OTexts, 2021, Chapter 6 "Judgmental forecasts"
  4. ↑ Rowe G., Wright G., The Delphi technique as a forecasting tool: issues and analysis, International Journal of Forecasting, 1999, pp. 353-375
  5. ↑ Gardner E.S., Exponential smoothing: The state of the art, Journal of Forecasting, 1985, pp. 1-28
  6. ↑ Hyndman R.J., Athanasopoulos G., Forecasting: Principles and Practice, 3rd edition, OTexts, 2021, section 7.6 "Forecasting with regression"
  7. ↑ Gneiting T., Katzfuss M., Probabilistic Forecasting, Annual Review of Statistics and Its Application, 2014, pp. 125-151
  8. ↑ Makridakis S., Spiliotis E., Assimakopoulos V., The M4 Competition: 100,000 time series and 61 forecasting methods, International Journal of Forecasting, 2020, pp. 54-74
  9. ↑ Hyndman R.J., Athanasopoulos G., Forecasting: Principles and Practice, 3rd edition, OTexts, 2021, section 11.2 "Single level approaches"
  10. ↑ Hyndman R.J., Koehler A.B., Another look at measures of forecast accuracy, International Journal of Forecasting, 2006, pp. 679-688
Types of forecasts — recommended articles
Forecasting — Moving average — Scenario analysis — Linear regression analysis — Sales planning — Strategic planning — Monte Carlo simulation
Retrieved from CEOpedia