Forecasting Foundations and Business Context

Introduction

Forecasting is the systematic process of using available information to form statements about uncertain future events. In business analytics, forecasting supports decisions such as how much inventory to purchase, how many employees to schedule, how much cash to reserve, and which markets to enter. A forecast is not a guarantee. It is an estimate conditional on information, assumptions, and a chosen method.

Forecasting differs from prediction in emphasis. The terms are often used interchangeably, but forecasting commonly refers to estimating future values of a variable over time, whereas prediction may refer more broadly to estimating an unknown outcome, including an outcome for a current or past observation. Forecasting is therefore especially concerned with temporal order, information availability, changing conditions, and the consequences of decisions made before the future is known.

A sound forecasting process begins with a business question rather than an algorithm. The analyst must clarify what is being forecast, for whom, at what frequency, over what horizon, and for what decision. A highly accurate forecast at the wrong level of aggregation or time horizon may be less useful than a moderately accurate forecast that fits the decision context.

Learning Objectives

After studying this chapter, you should be able to:

1. Define fundamental forecasting terminology and distinguish related concepts.
2. Explain forecasting horizons, time frequency, aggregation, and predictability.
3. Distinguish causal, extrapolative, judgmental, and quantitative approaches.
4. Match forecasting methods to business contexts and decision requirements.
5. Explain how forecasts are used in finance, marketing, operations, supply chains, workforce planning, and strategy.
6. Evaluate limitations, risks, ethical issues, and governance requirements in forecasting.
7. Design a basic forecasting problem statement and communicate its implications to decision-makers.

1. The Nature of a Forecasting Problem

A forecasting problem contains at least five elements:

• Target variable: the quantity or event to be estimated, such as weekly sales, monthly revenue, or daily demand.
• Forecast origin: the time at which the forecast is produced.
• Forecast horizon: how far into the future the forecast extends.
• Information set: data and knowledge available at the forecast origin.
• Decision context: the action that the forecast will inform.

Suppose a retailer produces a demand forecast on 1 June for each week of July. The forecast origin is 1 June, the target is weekly demand, and the horizon includes approximately four weekly periods. Information observed after 1 June must not be used to evaluate what could reasonably have been forecast on that date. This principle is essential: using future information creates look-ahead bias and gives a misleading impression of performance.

A time-series observation is commonly represented as y_t, where t identifies the time period and y is the measured variable. A forecast made at time t for a future period t+h can be written as:

Here, h is the forecast horizon and the vertical bar means “conditional on information available at.” The actual future value is y_t+h, while the forecast error is:

The error may be positive or negative. A positive error means that the actual value exceeded the forecast; a negative error means that the forecast was too high. Forecast error is not the same as a mistake in the ordinary sense: even a well-designed forecast can be wrong because the future contains genuinely unpredictable variation.

Forecasts, plans, targets, and scenarios

A forecast is an estimate of what is likely to happen. A plan describes what an organization intends to do. A target expresses a desired level of performance. A scenario is a coherent description of a possible future under specified assumptions. Confusing these concepts can produce poor decisions. For example, a sales target of $10 million is not evidence that sales will reach $10 million, and a budget may reflect management’s desired spending rather than an unbiased forecast of actual spending.

Organizations often use all four. A forecast may indicate expected demand of 90,000 units, a plan may authorize production of 85,000 units, a target may require 100,000 units, and scenarios may examine demand under weak, expected, and strong economic conditions.

2. Horizons, Frequency, and Aggregation

Forecast horizons

Forecast horizons are usually described as short, medium, or long term, but their meaning depends on the industry and decision cycle.

The same calendar period can have different meanings. A three-month forecast is long term for a daily trading desk but short term for a factory expansion decision. Forecast uncertainty generally increases as the horizon extends because more conditions can change. However, aggregation can sometimes reduce irregular short-term noise, making a monthly forecast easier than a daily forecast for some purposes.

Frequency and time intervals

Frequency is the interval at which observations are recorded: hourly, daily, weekly, monthly, quarterly, or annually. Frequency should be determined by the decision, data-generating process, and cost of measurement. A supermarket may need daily or hourly forecasts for staffing and replenishment, while a board may need quarterly or annual forecasts for strategic planning.

The frequency of data affects the patterns that can be detected. Daily data may contain day-of-week effects; monthly data may contain seasonal effects associated with holidays; annual data may hide those patterns. Changing frequency through aggregation can remove detail and may alter variability. Before aggregating, the analyst should determine whether the decision is based on totals, averages, end-of-period values, or another measure.

Lead time and decision timing

A forecast horizon should include operational lead time. If a supplier requires six weeks to deliver an item, a retailer cannot rely only on a one-week-ahead forecast when placing an order. The relevant forecast must cover the lead-time period and possibly a review period. In practice, forecasting and inventory policy are connected: the required horizon is determined partly by how quickly the organization can respond.

Cross-sectional and hierarchical structure

Forecasts may be required at several levels: total company sales, regional sales, store sales, product-category sales, and individual stock-keeping units. These forecasts form a hierarchy. If independently generated forecasts do not add up, managers face conflicting numbers. Forecast reconciliation is the process of making forecasts coherent across levels. It is not merely an arithmetic exercise; it requires decisions about which level is most reliable and where judgment should be applied.

3. Predictability and Sources of Variation

Predictability concerns the extent to which future values can be inferred from available information. It is not a fixed property of a variable. Predictability depends on the horizon, data quality, environment, measurement process, and information available.

A time series may contain several components:

• Level: the typical magnitude around which observations vary.
• Trend: a persistent upward or downward movement.
• Seasonality: a systematic pattern repeating at a known calendar frequency.
• Cycle: longer-term fluctuations whose length and timing are not necessarily fixed.
• Irregular variation: movement that is difficult to explain or forecast from available information.
• Structural change: a change in the process generating the observations.

A useful conceptual representation is:

The signal includes systematic information that may be forecast, while noise represents variation not reliably predictable with the available information. This is a conceptual decomposition rather than a universal statistical model. In some applications, a multiplicative representation is more appropriate, especially when seasonal variation grows with the level:

The additive form is appropriate when the magnitude of seasonal fluctuations is roughly constant; the multiplicative form is more appropriate when fluctuations are proportional to the level and values are positive. Analysts should not select a form solely because it appears mathematically convenient.

Signal, noise, and the forecast ceiling

If demand is affected by unobserved customer preferences, random weather, competitor actions, or recording errors, perfect forecasting may be impossible. The practical objective is not to eliminate all error but to reduce avoidable error, quantify remaining uncertainty, and make decisions robust to it.

Forecastability can be improved by:

• collecting timely and relevant data;
• distinguishing measurement error from genuine variation;
• incorporating known events such as promotions or holidays;
• modeling structural breaks explicitly;
• using an appropriate level of aggregation; and
• combining statistical evidence with carefully governed expert knowledge.

Yet additional complexity does not guarantee improvement. A complex model can overfit historical noise and perform poorly on new data. Overfitting occurs when a method captures accidental features of the historical sample rather than general patterns. Evaluation on observations not used to fit the model is therefore essential.

4. Causal and Extrapolative Forecasting

Extrapolative forecasting

Extrapolative, or univariate, forecasting uses the historical behavior of the target variable itself. Examples include naive forecasts, moving averages, exponential smoothing, and autoregressive models. These methods are useful when the series contains stable temporal patterns and relevant explanatory variables are unavailable or unreliable.

A simple naive forecast sets the next forecast equal to the most recent observation:

This method is a strong baseline for many series and is particularly useful when the process follows a random-walk-like pattern. A seasonal naive forecast uses the value from the corresponding previous season:

where m is the number of observations in one seasonal cycle, such as 12 for monthly data with annual seasonality. These formulas are simple, but their applicability depends on the relevant persistence or seasonal repetition.

A moving-average forecast for a one-step-ahead value may be written as:

where k is the number of recent observations included. The method gives equal weight to those observations and assumes that a recent average is informative about the next period. Larger k produces more smoothing but may respond more slowly to changes.

Causal forecasting

Causal forecasting relates the target to explanatory variables. A basic linear relationship can be written as:

Here, y_t is the target, x_j,t are explanatory variables, β₀ is the intercept, β_j are coefficients, and ε_t is an error term. The equation describes an association unless a credible design and assumptions justify a causal interpretation. A coefficient does not automatically mean that changing x_j will cause the predicted change in y.

Causal or explanatory models are valuable when known drivers are expected to change. For example, a retailer may forecast sales using price, advertising expenditure, store traffic, holidays, and competitor activity. If management plans a price change, a model containing price may be more useful than a purely extrapolative model.

Causal models require care. Explanatory variables must be available, or themselves forecastable, at the time the target forecast is made. Relationships can change, variables may be correlated, and omitted factors can distort interpretation. Correlation is not proof of causation, and a model that fits historical data may fail when policy, customer behavior, or market conditions change.

Choosing between approaches

The choice is not always either causal or extrapolative. Hybrid systems may use a time-series model for residual patterns after accounting for promotions, or combine expert judgment with statistical forecasts. The appropriate approach depends on:

1. the purpose of the forecast;
2. the stability of historical patterns;
3. availability and quality of drivers;
4. the expected effect of future interventions;
5. required transparency; and
6. the cost of errors and model maintenance.

5. Judgmental and Quantitative Forecasting

Judgmental forecasting

Judgmental forecasting relies primarily on informed human assessment. It is appropriate when historical data are scarce, the environment is undergoing structural change, or the forecast concerns a novel product or strategic event. Common methods include expert panels, sales-force estimates, management consensus, analogy, and structured techniques such as the Delphi method.

Judgment can provide information unavailable in a database, including knowledge of pending competitor actions, customer negotiations, regulatory developments, or supply disruptions. However, unaided judgment is vulnerable to anchoring, overconfidence, availability bias, group pressure, and incentives to produce favorable numbers.

Good practice separates the statistical baseline from the judgmental adjustment. The analyst should record who made the adjustment, its magnitude, its rationale, and the subsequent outcome. This creates accountability and permits learning.

Quantitative forecasting

Quantitative methods use numerical data and explicit procedures. They include time-series models, regression, simulation, machine learning, and probabilistic models. Their strengths include reproducibility, scalability, and the ability to evaluate performance systematically. Their weaknesses include sensitivity to data quality, structural change, and inappropriate assumptions.

A quantitative forecast should be accompanied by uncertainty information where possible. A point forecast gives one central estimate; an interval forecast gives a range intended to contain the future value with a stated probability under model assumptions. Wider intervals generally indicate greater uncertainty, but interval quality must be assessed empirically rather than inferred from width alone.

Combining methods

Combining forecasts can improve robustness when methods contain different information or make different errors. A simple combination may average two forecasts:

where 0 ≤ w ≤ 1 is the weight assigned to method 1. The weights should be selected using an appropriate validation procedure or justified business knowledge. Combining is not automatically superior, but it can reduce dependence on a single model.

6. Forecasting Across Business Functions

Finance and accounting

Financial forecasting supports cash-flow planning, revenue and expense estimation, credit risk assessment, liquidity management, and valuation. Treasury teams may forecast daily cash balances, while corporate finance may forecast quarterly earnings. Forecasts affect borrowing, investment, covenant compliance, and investor communication.

Financial forecasts are exposed to interest-rate changes, exchange rates, inflation, credit conditions, and market shocks. Governance is particularly important because biased forecasts can influence reported expectations and capital allocation. Analysts should distinguish operational forecasts from targets and document assumptions about prices, volumes, timing, and foreign exchange.

Marketing and customer analytics

Marketing forecasts estimate demand, customer acquisition, conversion, churn, campaign response, and customer lifetime value. Forecasts can guide media allocation and promotion timing. A campaign may increase observed sales, but analysts must consider seasonality, concurrent campaigns, selection effects, and sales that would have occurred without the campaign. Causal experiments or carefully designed quasi-experiments are stronger than simple before-and-after comparisons when the goal is to estimate campaign impact.

Operations

Operations forecasting supports production scheduling, capacity planning, maintenance, service staffing, and quality management. Forecasts may concern units required, machine failures, processing times, or service arrivals. Errors can create either excess capacity and cost or shortages and delays. The relevant forecast is often not only the expected value but also the distribution of possible demand, because capacity and safety decisions depend on risk.

Supply chains and inventory

Supply-chain forecasting connects customer demand with procurement, production, transportation, and inventory. Forecasts must account for lead times, order cycles, promotions, substitutions, stockouts, and information delays. A low recorded sales figure may reflect low demand—or simply that the product was unavailable. Treating censored sales as true demand can systematically underforecast future requirements.

Forecasts should be coordinated across the supply network where appropriate. The bullwhip effect describes the amplification of demand variability as information moves upstream from customers to suppliers. Delayed information, batch ordering, price promotions, and rationing can contribute to this effect. Better information sharing and coordinated planning can reduce, though not necessarily eliminate, amplification.

Workforce planning

Organizations forecast headcount, hiring needs, absenteeism, turnover, workloads, and skill requirements. Workforce forecasts must address legal and ethical concerns because predictions about individuals can affect employment opportunities. Models should be tested for disparate performance across relevant groups, and high-impact decisions should not rely on opaque automated scores alone.

Strategy and investment

Strategic forecasting informs market size, technology adoption, facility location, competitive dynamics, and investment timing. Long-term forecasts are especially vulnerable to structural breaks and uncertainty. Scenarios, sensitivity analysis, and real-options thinking can complement point forecasts. Strategic decisions should not be based on a single precise number when the underlying uncertainty is substantial.

7. A Professional Forecasting Workflow

A disciplined workflow is more important than selecting a fashionable algorithm.

1. Define the decision. Identify the action, decision-maker, cost of delay, and consequences of over- or underforecasting.
2. Specify the target. Define the variable, unit, population, location, horizon, frequency, and aggregation level.
3. Establish the information set. Record what was known at each forecast origin and prevent leakage from the future.
4. Explore the data. Examine missing values, outliers, changes in measurement, seasonality, trend, promotions, and stockouts.
5. Create benchmarks. Use simple methods such as naive or seasonal naive forecasts before evaluating complex models.
6. Select candidate methods. Match the method to data properties, explanatory information, interpretability, and operational constraints.
7. Validate through time. Use rolling-origin or time-based evaluation rather than randomly shuffling observations when temporal order matters.
8. Evaluate relevant metrics. Select measures aligned with the decision and communicate their limitations.
9. Generate and communicate uncertainty. Explain assumptions, intervals, scenarios, and important risks.
10. Monitor and revise. Track forecast errors, data changes, bias, drift, and the effects of interventions.

Common accuracy measures include mean absolute error:

and root mean squared error:

Here, n is the number of evaluated forecasts, y_i is an actual value, and ŷ_i is its forecast. Both measures have the same units as the target. RMSE penalizes large errors more heavily than MAE because errors are squared before averaging. Neither measure alone identifies whether errors are systematically too high or too low.

Percentage measures can be problematic when actual values are zero or close to zero. Metrics should therefore be selected with awareness of scale, intermittency, business cost, and the consequences of asymmetric errors. A forecast can have good average accuracy while systematically underforecasting peak demand.

8. Illustrative Case: Forecasting Demand for a Meal-Delivery Service

Background

The following is a fictional illustrative case. FreshRoute delivers prepared meals to customers in three metropolitan areas. It must decide weekly ingredient purchases and driver capacity. Management currently uses last month’s average demand, but frequent promotions and holiday effects produce shortages on some days and food waste on others.

Problem

The operations manager asks for a forecast of next week’s orders by city and meal category. The decision requires forecasts seven days ahead, while some ingredients have supplier lead times of five days. The data include historical orders, promotions, holidays, weather indicators, cancellations, and stockout records.

Analysis

The analyst first clarifies that recorded orders are not always equal to unconstrained demand. On days when a meal was unavailable, observed orders may understate the number customers would have purchased. The analyst therefore flags stockout periods rather than treating them as ordinary low-demand observations.

A seasonal-naive benchmark is created using demand from the same weekday in the previous week. A causal candidate model includes promotion status, holiday indicators, city, meal category, and weather information available at the forecast origin. A judgmental review is then conducted with the marketing and operations teams, who know that a large promotion has been approved but not yet fully reflected in historical data.

The team evaluates forecasts using rolling weekly origins. MAE is reported in orders, because the operations manager understands this unit. Errors are also examined separately for promoted and non-promoted days. This prevents a strong average result from hiding poor performance during the events that matter most.

Approach and solution

The final process uses a statistical baseline plus a documented promotional adjustment. The baseline is retained for comparison, and the adjustment is limited to evidence supported by previous promotion response and the planned campaign size. The output includes a central forecast and a high-demand scenario for driver scheduling. Procurement receives forecasts at meal-category level, while senior management receives an aggregated city-level view.

Outcome

In the first evaluation cycle, the combined approach reduces typical error relative to the old monthly-average method and makes promotional assumptions visible. It does not remove uncertainty: an unexpected heatwave changes customer preferences, and one supplier experiences a delay. Because the process includes a high-demand scenario and supplier communication, the operational impact is smaller than it would have been under a single point forecast.

Learning points

The case illustrates that forecasting quality depends on problem definition, data interpretation, benchmarks, validation, organizational knowledge, and communication. A sophisticated model cannot correct for stockouts recorded as demand, an incorrect horizon, or an undocumented management adjustment.

9. Limitations, Ethics, and Governance

Forecasting systems can reproduce historical inequities, amplify misleading measurements, or encourage false precision. Ethical practice requires attention to the purpose of the forecast, affected stakeholders, data provenance, privacy, fairness, explainability, and accountability.

Important governance practices include:

• documenting data sources, transformations, model versions, and assumptions;
• restricting access to personal and commercially sensitive data;
• testing performance across relevant groups and operating conditions;
• monitoring drift and structural change;
• preserving human review for high-impact decisions;
• distinguishing forecast uncertainty from managerial preference; and
• maintaining an audit trail for overrides and consequential decisions.

AI and machine-learning systems can process large numbers of predictors and detect nonlinear patterns, but they do not eliminate the need for domain knowledge. They may be especially vulnerable to distribution shift: performance deteriorates when future conditions differ materially from training data. Automated systems should therefore be benchmarked against simple methods, evaluated using realistic time-based tests, and monitored after deployment.

Forecast communication is also a governance issue. A point estimate such as 50,000 units may be interpreted as certainty unless its assumptions and uncertainty are explained. Decision-makers should know whether the number is a baseline forecast, a target, a scenario, or a judgmentally adjusted value.

10. Best Practices for Beginning Analysts

A beginning analyst should:

1. Start with the decision and required horizon.
2. Define the target precisely and confirm its units.
3. Check whether the data measure demand, sales, capacity, or an outcome constrained by availability.
4. Establish simple benchmarks before using complex models.
5. Preserve temporal order during validation.
6. Avoid future information leakage.
7. Separate forecasts from targets and plans.
8. Report bias and uncertainty, not only average accuracy.
9. Document judgmental adjustments.
10. Reassess the process when the environment, policy, product, or measurement system changes.

Chapter Summary

Forecasting is the disciplined estimation of uncertain future outcomes for a defined decision context. A forecasting problem requires a target, forecast origin, horizon, information set, frequency, and level of aggregation. Forecasts differ from plans, targets, and scenarios, although organizations commonly use all of them together.

Predictability depends on systematic signal, random variation, data quality, horizon, and structural stability. Time-series behavior may include level, trend, seasonality, cycles, irregular variation, and structural change. Extrapolative methods use the history of the target, while causal methods incorporate explanatory variables; neither is universally superior. Judgmental methods are valuable when data are scarce or conditions are changing, but they must be structured and documented. Quantitative methods support reproducibility and evaluation, but they remain dependent on assumptions and data quality.

Forecasting supports finance, marketing, operations, supply chains, workforce planning, and strategy. Professional practice includes clear problem definition, benchmark models, time-based validation, appropriate error measures, uncertainty communication, monitoring, and governance. Ethical forecasting requires privacy protection, fairness assessment, transparency, accountability, and caution against false precision. The central professional lesson is that forecasting is not merely a modeling task: it is an evidence-based decision process under uncertainty.

Key Terms

Forecast: An estimate of a future value or event based on available information.

Forecast origin: The time at which a forecast is produced.

Forecast horizon: The distance into the future covered by a forecast.

Forecast error: The difference between an actual outcome and its forecast.

Time series: Ordered observations measured over time.

Frequency: The interval between successive observations.

Aggregation: Combining observations across time, products, locations, or organizational levels.

Predictability: The extent to which future outcomes can be inferred from available information.

Trend: A persistent long-term upward or downward movement.

Seasonality: A repeating pattern associated with a known calendar frequency.

Cycle: A longer-term fluctuation whose timing and duration may vary.

Extrapolative forecasting: Forecasting based primarily on the historical behavior of the target series.

Causal forecasting: Forecasting that uses explanatory variables related to the target.

Judgmental forecasting: Forecasting based primarily on expert or managerial assessment.

Quantitative forecasting: Forecasting based on numerical data and explicit analytical procedures.

Baseline: A simple reference method against which more complex forecasts are compared.

Overfitting: Modeling historical noise rather than generalizable patterns.

Look-ahead bias: Using information that would not have been available when the forecast was made.

Forecast reconciliation: Adjusting forecasts so that values at different hierarchical levels are coherent.

Bullwhip effect: Amplification of demand variability as information moves upstream through a supply chain.

Review Questions

5 Short Answer Questions

1. What are the forecast origin and forecast horizon?
2. How does a forecast differ from a target and a plan?
3. Define seasonality and provide one business example.
4. Why is a naive forecast useful even when more advanced methods are available?
5. What is look-ahead bias, and why does it damage forecast evaluation?

5 Long Answer Questions

1. Explain the differences between extrapolative and causal forecasting, including the circumstances in which each may be appropriate.
2. Discuss how forecast frequency, aggregation, and lead time influence a business forecasting problem.
3. Compare judgmental and quantitative forecasting, including their principal strengths, risks, and governance requirements.
4. Explain how forecasting supports finance, marketing, operations, supply chains, workforce planning, and strategy.
5. Describe a professional forecasting workflow from problem definition through monitoring and revision.

3 Analytical / Application Questions

1. A retailer has daily sales data, frequent stockouts, and a six-week supplier lead time. Design a forecasting problem statement and identify the data issues that must be addressed before modeling.
2. A marketing team wants to claim that a campaign caused a sales increase because sales were higher after the campaign than before it. Analyze why this comparison may be insufficient and propose a stronger evaluation approach.
3. A company’s monthly demand forecast is accurate on average but consistently underestimates demand during holidays. Recommend changes to the forecasting process, validation design, performance measures, and communication of uncertainty.

Multiple Choice Questions with Answers

Question 1

Which statement best defines a forecast?

A. A desired future performance level

B. An estimate of a future outcome based on available information

C. A guarantee that a future event will occur

D. A historical record of business activity

Correct Answer: B. An estimate of a future outcome based on available information

Question 2

In the notation ŷ_t+h|t, what does h represent?

A. The forecast error

B. The number of explanatory variables

C. The forecast horizon

D. The historical average

Correct Answer: C. The forecast horizon

Question 3

Which method uses the historical behavior of the target variable without requiring explanatory variables?

A. Extrapolative forecasting

B. Causal forecasting

C. Experimental design

D. Scenario planning only

Correct Answer: A. Extrapolative forecasting

Question 4

A seasonal-naive forecast for monthly data with annual seasonality would generally use which observation for forecasting next January?

A. The previous February

B. The previous January

C. The average of all Januaries and Februaries

D. The current December only

Correct Answer: B. The previous January

Question 5

Why can a stockout create a forecasting problem?

A. It always increases recorded demand

B. It may cause recorded sales to understate unconstrained demand

C. It removes all seasonality from the data

D. It guarantees that future demand will be zero

Correct Answer: B. It may cause recorded sales to understate unconstrained demand

Question 6

Which metric penalizes large errors more heavily because errors are squared before averaging?

A. MAE

B. RMSE

C. Median absolute error

D. Forecast bias only

Correct Answer: B. RMSE

Question 7

What is look-ahead bias?

A. Using a longer forecast horizon than intended

B. Using information that would not have been available when the forecast was made

C. Forecasting a seasonal series

D. Comparing two forecasts with different units

Correct Answer: B. Using information that would not have been available when the forecast was made

Question 8

Which statement about a coefficient in a causal forecasting model is most accurate?

A. It automatically proves causation

B. It is always measured in percentages

C. It describes an association unless causal assumptions and design justify a causal interpretation

D. It cannot change when the environment changes

Correct Answer: C. It describes an association unless causal assumptions and design justify a causal interpretation

Question 9

What is the primary purpose of forecast reconciliation?

A. To remove all forecast uncertainty

B. To make forecasts at different hierarchical levels add up coherently

C. To replace quantitative models with judgment

D. To increase the forecast horizon automatically

Correct Answer: B. To make forecasts at different hierarchical levels add up coherently

Question 10

Which practice is most appropriate when validating a time-series forecast?

A. Randomly shuffle all observations before splitting the data

B. Use future observations to tune the historical model

C. Preserve temporal order and evaluate using time-based or rolling-origin tests

D. Evaluate only the largest observed values

Correct Answer: C. Preserve temporal order and evaluate using time-based or rolling-origin tests