When discussing forecasts, the focus often defaults to accuracy metrics like MAPE and MSE. Teams compare these figures, select the model with the lowest number, and declare it the winner. However, relying solely on this method is not always the best approach.
Choosing a forecast model is also about “risk”: what kind of mistakes you are willing to accept, and which ones you want to avoid. Some models are more “speculative”: they react quickly to recent changes and give accurate short-term results. However, they sometimes miss badly. Other models are more “conservative”: they change slowly, smooth the data, and avoid extreme outcomes. However, they might not achieve the smallest average error.
Neither is always better, as the right choice depends on decision risk.
Two types of forecast risk
1. Volatility risk: error spread
Some models produce many accurate predictions and rare but large misses (Figure 1)
Figure 1: Typical errors in a “speculative” model
Others produce steady moderate errors, but very few extremes (Figure 2)
Figure 2: Typical errors in a more “conservative” model
Even when two models have similar average accuracy, their operational value can differ. In some situations, large errors carry disproportionately higher costs than small ones. In regulated service lines (e.g., insurance industry), a big miss can trigger formal complaints and audit risk.
2. Level risk: reacting too much to recent change
Models also differ in how strongly they trust the latest data.
When volumes rise suddenly:
- A responsive model assumes the new level will continue (Figure 3)
- A more “stable” model assumes partial return to normal (Figure 4)
If forecasts are refreshed often, a fast reaction is useful. If forecasts are refreshed rarely, overreaction is risky, as you may lock in the wrong level for too long.
Figure 3: Responsive model
Figure 4: A “stable” model
A simple analogy: the fever forecast
Consider a medical example.
If a patient has a fever of 40°C today:
- A good forecast for tomorrow is close to 40°C
- A good forecast for next year is not 40°C
Short-term forecasts should follow the current signal, whereas long-term forecasts should assume recovery and return to normal.
In many contact centers, short-term spikes tend to move back toward “normal” levels over time (in statistics, this is typically called “mean reversion”). However, not all data behaves this way. Some processes show a clear and lasting trend instead.
When demand displays a mean reversion:
- Operational forecasts should follow recent level shifts
- Strategic forecasts should discount sudden spikes
- Tactical forecasts should react, but with limits
The difference is about how long you expect changes to last.
Operational forecasts: small daily / interval errors matter most
Short-term operational forecasts support scheduling and near-term staffing. They are refreshed frequently, and corrective actions are available.
A simple operational decision tree might look like this (numbers are just illustrative*):
Error Range | Decision Impact |
|---|---|
Under 5% | No action needed |
5–20% | Manageable cost impact |
Above 20% | Action needed |
Mitigation can include overtime, shift moves, backlog handling, and short-term adjustments
Because forecasts are rerun often, small daily errors create steady savings, whereas occasional large misses can be handled.
Here, a more speculative model, with fast reactions, is typically more suitable.
*In practice, decision impact is driven by absolute error and cost sensitivity, not percentage error.
Strategic forecasts: large level errors matter most
Strategic forecasts support hiring, budgeting, and long-term capacity decisions. They are refreshed infrequently (one or twice per year), and corrections are slow.
A simple strategic decision rule might be:
- Within ~20% → staffing acceptable
- Beyond ~20% → hire or reduce staff
There is no quick correction cycle. If the model overreacts to a temporary spike, the organisation may hire too many people and carry excess capacity for months.
The main danger is not daily error spread. It is choosing the wrong level and keeping it for too long.
Here:
- Stability is more important than fast reaction
- Avoiding extremes matters more than daily accuracy
- Conservative behaviour reduces long-term risk
Tactical forecasts: something in the middle
Mid-term planning sits between operational and strategic horizons.
The aim is controlled reaction:
- Respond to real change
- Avoid short-term noise
- Limit large swings
Typical controls include a moderate smoothing or test scenarios generation.
When business knowledge overrides the model
Sometimes you know that a change is permanent.
Examples:
- A new contract is signed
- A product line is closed
- A channel is migrated
- Pricing or policy has structurally changed
In these cases, relying purely on model behaviour, whether speculative or conservative, is not enough. You adjust the forecast based on business knowledge and expectations.
That introduces subjectivity and should be made explicit:
- Document the assumption
- Quantify the adjustment
- Separate model error from judgment error
Good forecasting practice combines models and judgement, with clear visibility of where human assumptions enter the forecast.
Practical rule
In conclusion, do not ask only: “Which model is most accurate?”
Also ask:
- How costly are large misses?
- How often will this forecast be refreshed?
- How easy is it to reverse the decision?
- Am I adding business judgment?
Forecast maturity is about choosing the right risk behaviour for each planning horizon and applying judgment in a controlled way.
Technical disclaimer
Since this is CCmath, and precision matters here, a technical disclaimer is added.
The examples in this article are illustrative. Actual forecast error behaviour depends on the data, model structure, parameter tuning, and the loss function used. In theory, risk preference is built into the loss function, so if decision costs are known and measurable, choosing the right loss function already reflects risk in model selection.
In practice, however, the “correct” loss functions are difficult to define and maintain across all contact center lines and skills (which can theoretically be different per line/skill). Teams, therefore, rely on shared proxy metrics such as MAPE or WAPE. Optimising only these metrics does not always lead to the best operational decisions.
The terms “speculative” and “conservative” are used here as practical descriptions of model behaviour under different parameter settings, not as formal statistical categories.
From forecast risk to forecast control
At CCmath, this risk trade-off is made explicit in our forecasting software, CCforecast. The platform lets you control responsiveness and stability through model parameter settings, with separate configurations for long-term, tactical, and short-term forecasting horizons. Book a demo today to learn more.