What are the challenges in forecasting? 8 problems and fixes

The main challenges in forecasting are poor or inconsistent data, the future behaving differently from the past, structural breaks such as a new competitor or a pandemic, bias in human adjustments, confusing forecasts with targets, using the wrong accuracy metric, and teams working in silos. Most can be reduced with clean data, simple combined models and honest error tracking.

Key takeaways

• How predictable something is depends on four things: how well you understand its drivers, how much data you have, how similar the future is to the past, and whether the forecast changes the outcome.
• Data problems (missing history, changed definitions, outliers, stock-outs recorded as low demand) cause more forecast error than the choice of model.
• A forecast is not a goal. Mixing the two produces optimistic numbers that nobody trusts.
• Averaging the output of a few simple methods is one of the most reliable ways to improve accuracy.
• Override statistical forecasts only when you have real information the model lacks, and log every override so you can measure whether it helped.
• Track error with a metric that suits your data. MAPE breaks down when actual values are zero or close to zero.

Why is forecasting difficult?

Forecasting is hard because you predict an outcome from incomplete information, and the conditions that produced your history keep changing. Some quantities are easy to forecast and some are close to impossible.

The textbook Forecasting: Principles and Practice by Rob Hyndman and George Athanasopoulos gives a useful test. A quantity is predictable when you understand what drives it, have enough data, can assume the future resembles the past, and the forecast itself does not change behaviour. Short-term household electricity demand meets all four conditions, so it can be forecast well. Currency exchange rates meet only one: there is plenty of data, but the drivers are poorly understood and markets react to the forecasts themselves.

Most business forecasts sit between those extremes. Next month's support ticket volume is fairly predictable. Sales of a product you have never sold before are not. Knowing where your forecast sits on that scale tells you how much effort and how much trust it deserves. For a longer discussion of the root causes, see our article on why forecasting is difficult.

The eight most common forecasting challenges

1. Poor and inconsistent data

Every forecast inherits the quality of its input. Typical faults include missing periods, product codes that changed during an ERP migration, duplicated orders, and returns booked against the wrong month.

A subtler problem is recording sales instead of demand. If a product was out of stock for three weeks, sales were low because of supply, not because customers stopped wanting it. A model trained on those weeks will under-forecast the next season. Fix this by flagging stock-outs, promotions and one-off events in the data before modelling, and by agreeing one definition of "demand" across sales, finance and supply chain.

2. Too little history, or too much of the wrong kind

New products, new markets and new customer segments have no history at all. In those cases judgement and analogies with similar products are the only option, and the forecast should carry a wide range.

The opposite problem also exists. Ten years of history may include a business that no longer exists in the same form. Old data from a different product mix, price level or channel can pull a model in the wrong direction. Use the history that still describes how the business works today.

3. Structural breaks and external shocks

A new competitor, a regulation, a pandemic or a sudden change in interest rates can break the link between past and future. No model trained only on history will see these coming.

You cannot remove this risk, but you can limit the damage. Run scenarios for the shocks you can name, shorten the planning cycle so forecasts are refreshed when conditions change, and watch leading indicators such as order intake, quote volume or web enquiries rather than waiting for booked revenue.

4. Seasonality and calendar effects

Weekly patterns, holidays that move between months, the number of working days and long seasonal cycles all distort simple models. Weekly data is awkward because a year holds about 52.18 weeks, not a whole number. Statistical agencies deal with this using dedicated seasonal adjustment tools such as the US Census Bureau's X-13ARIMA-SEATS. Most forecasting libraries now include decomposition methods that separate trend, season and remainder, so calendar effects can be modelled explicitly instead of being treated as noise.

5. Confusing forecasts with targets

Hyndman and Athanasopoulos separate three ideas. A forecast predicts what will happen given everything you know. A goal is what you would like to happen. A plan is what you will do to close the gap between the two.

When the sales forecast doubles as the sales target, people push it up to look ambitious or down to make it easy to beat. Either way it stops being a forecast. Keep the statistical forecast, the target and the plan as separate numbers, owned by different people if possible.

6. Bias in human overrides

Managers often adjust statistical forecasts, and many adjustments make them worse. The chapter on judgmental forecasts in the same textbook summarises the research: practitioners adjust far more often than they should, small positive adjustments tend to hurt accuracy, and adjustments help most when they are based on information the model does not have, such as a confirmed large order or a planned price change.

A practical rule: require a written reason for every override, store the original and adjusted forecast side by side, and review after the fact which overrides improved the result.

7. Choosing the wrong model, or only one model

Teams often chase the most sophisticated algorithm. In practice, simple methods such as exponential smoothing or ARIMA are hard to beat on many business series, and combining several of them is better still. The textbook's section on forecast combinations notes that research since Bates and Granger in 1969 has found that averaging forecasts from different methods usually improves accuracy, and that a simple average is hard to beat with more complex weighting.

Machine learning helps when you have many related series and useful external drivers, such as weather, prices or web traffic. It adds less when you have one short series and nothing else. Our guide to the best model for sales forecasting compares the main options.

8. Measuring accuracy badly, or not at all

Many organisations never compare forecasts with actuals in a structured way, so they cannot tell whether forecasting is improving. Others use a metric that misleads. Mean absolute percentage error (MAPE) is popular, but it is undefined when an actual value is zero and becomes extreme when actuals are close to zero, as the textbook's section on evaluating forecast accuracy explains. That makes it a poor choice for slow-moving items or intermittent demand.

Pick a metric that suits the data, such as mean absolute error or a scaled error, test models on a hold-out period they were not trained on, and always compare against a naive benchmark like "same as last period". If your model cannot beat the naive forecast, it is not adding value.

How do the challenges compare?

How do you build a forecasting process that holds up?

A reliable process matters more than any single tool. These steps work for demand, revenue and capacity forecasts alike:

1. Define the decision. State what the forecast is for, the time horizon and the level of detail, such as weekly units per warehouse for the next 12 weeks.
2. Agree one data source. Pull history from one governed source and document known gaps and events.
3. Start with benchmarks. Run naive and seasonal naive forecasts first, then add models only if they beat them.
4. Combine methods. Average two or three reasonable models rather than betting on one.
5. Bring in the business. Hold a short monthly review where sales, finance and operations add known information, with every override logged.
6. Measure and publish error. Track accuracy and bias per product group and share it, so trust is based on evidence.

Tooling then supports the process: a data platform that keeps history clean, automated model runs and dashboards that show forecast against actual. Our overview of demand forecasting tools covers the common options.

A cloud example: forecasting seasonal capacity

Forecasting challenges also appear in IT. Capacity planning is a forecast of load, and a wrong forecast means either slow systems or paid-for servers sitting idle.

Branäsgruppen, one of Sweden's largest ski resort groups, makes most of its trade between December and April and reaches 15,000 riders per day at peak. On its old VMware environment, web front-ends and databases had to be scaled by hand ahead of demand, and performance dropped before each scaling event. After Opsio migrated the estate to AWS with auto-scaling across multiple Availability Zones, capacity follows demand automatically. The case reports 40% faster average page responsiveness and a 40% reduction in total cost of ownership. The forecast still matters for budgeting, but a missed peak no longer takes the website down.

How Opsio helps

Opsio builds the data and cloud foundations that forecasting depends on: governed data pipelines, model hosting and monitoring, and dashboards that compare forecast with actual. Our predictive analytics consulting team helps you pick sensible methods and measure them honestly, and our data and AI services cover the platform underneath.

Frequently asked questions

What are the main challenges of forecasting?

The main challenges are poor data quality, too little relevant history, structural breaks and external shocks, seasonality, using forecasts as targets, biased manual overrides, over-reliance on one complex model, and weak accuracy tracking. Data and process problems usually cause more error than the choice of algorithm.

What are the advantages and disadvantages of forecasting?

Forecasting helps you plan staffing, inventory, budgets and capacity before demand arrives. The disadvantages are that every forecast carries uncertainty, it can create false confidence if presented as a single number, and it takes ongoing effort to maintain data and models. Showing a range rather than a point estimate reduces the risk of false confidence.

How can forecast accuracy be improved?

Clean the history, compare every model against a naive benchmark, combine a few simple methods, limit manual overrides to cases with real new information, and measure error every cycle with a metric that suits your data.

Why do sales forecasts often miss?

Sales forecasts often double as targets, so they absorb optimism or sandbagging. They also rely on pipeline data that depends on how each salesperson records deals. Separating the forecast from the quota and using historical conversion rates by stage makes them more reliable.

Can AI solve forecasting challenges?

AI and machine learning help when you have many related series and good external drivers, and they automate model selection at scale. They do not fix bad data, structural breaks or forecasts used as targets. Treat them as one more method to combine and test, not a replacement for the process.