Time Series Forecasting for Finance Teams

Time series forecasting for finance teams uses the pattern in a variable's own history, such as monthly revenue, weekly cash receipts or quarterly expenses, to project its future values. The idea is simple: if sales have grown steadily and peaked every December, the best first guess for next year is another year of growth with a December peak. Time series methods are fast, cheap and surprisingly hard to beat for stable series. They fail when the world changes in ways the history does not show. This guide explains the components of a time series, shows a complete worked example of calculating seasonal indexes and a forecast, covers the steps for testing a model and sets out when to switch to other methods. It is part of the forecasting section of budgeting and forecasting and extends financial forecasting.

The components of a time series

Any series of data over time can be thought of as a combination of parts.

Forecasting aims to estimate level, trend and seasonality, and to treat noise as unpredictable. Understanding which components exist in your data guides the method.

Step 1: Look at the data

Plot the series. Look for:

• A trend, up or down.
• Repeating peaks and troughs at the same time of year.
• Breaks: a sudden change in level because of a price change, acquisition or lost customer.
• Outliers: one-off events that should be adjusted.
• Missing data or changes in definitions.

Do not skip this step. A model fitted to uncorrected data will project the problems forward.

Step 2: Clean and adjust

• Remove or adjust one-off items (a settlement, a pull-forward order).
• Adjust for known price changes or changes in the number of working days.
• Use consistent definitions across the whole history.
• Keep at least two full seasonal cycles if you want to estimate seasonality (24 months for monthly data).

Step 3: Estimate seasonality

A seasonal index shows how a period compares with the average. Here is a complete example with quarterly data (in thousands of dollars).

For each year, divide each quarter by that year's average, then average across years.

The indexes sum to 4.00 (0.8 + 1.0 + 0.9 + 1.3), which is the check: indexes for four quarters must average 1.0. This illustrative series is very regular; real data will vary from year to year, so average several years where available.

Step 4: Estimate the level and trend

The yearly average rose from 100 to 110, growth of 10 percent. If that growth continues, the average for year 3 is 110 × 1.10 = 121.

Step 5: Build the forecast

Forecast = Baseline level for the period × Seasonal index.

Check: 4 × 121 = 484. The forecast preserves the annual total implied by the trend and distributes it with the seasonal pattern. This is the same logic used in budgeting to phase an annual target by month. See adjusting forecasts for seasonality.

A refinement: in a growing series, the baseline rises through the year, so quarter-by-quarter growth could be applied to the baseline as well. With 10 percent annual growth, the baseline in each quarter is about 2.4 percent above the previous quarter, since 1.10 to the power of one quarter is 1.0241. Whether to include the within-year ramp depends on the pattern in your data.

Step 6: Choose a method

Start with the simplest method that fits the pattern. Move to more complex ones only if they win a fair test.

Step 7: Test the model

1. Hold out recent data. Fit the model on all but the last four to six periods, and forecast those.
2. Compare with a naive benchmark. If you cannot beat last year's same month, the model is not adding value.
3. Measure accuracy and bias. Use mean absolute error and mean absolute percentage error, and check the average error for bias. See how to measure financial forecast accuracy.
4. Look at the errors. Do they show a pattern, such as under-forecasting every Q4?
5. Check reasonableness. Does the forecast make sense given what you know about the business?

Step 8: Use it

Publish the forecast with a range, document the method and adjustments, and track accuracy every period.

When time series methods work, and when they do not

For intermittent items or lumpy project revenue, driver-based or pipeline methods are better. See driver-based forecasting and pipeline-based sales forecasting.

Applying it in finance

Time series in product businesses

For product companies, statistical baselines by item and location are the starting point of the demand plan. Planners then overlay known events. Accuracy at that level affects safety stock and cash. See demand planning and inventory optimization.

Linking to unit economics

A time series forecast tells you how many units; economics tell you what they are worth. Combine the two to forecast profit, not just volume. See unit economics.

Common mistakes

• Ignoring structural breaks and fitting through them.
• Using too little history for seasonality.
• Not adjusting for one-offs.
• Choosing a complex model without testing against a simple one.
• Forecasting far beyond the stable pattern.
• Not tracking accuracy and bias.

Related guides

For a comparison of two of the most common methods, read moving average vs. exponential smoothing for forecasting. For adding drivers to statistics, see regression analysis for financial forecasting in Excel.

FAQs

What is time series forecasting?

Forecasting a variable using patterns in its own past values, such as trend and seasonality.

How much data do I need for time series forecasting?

At least two full seasonal cycles for seasonality (24 months for monthly data), and more is better.

What is a seasonal index?

A ratio showing how a period compares with the average, such as 1.30 for a quarter that is typically 30 percent above average.

How do I know if my forecast model is good?

Test it on held-out data, compare with a naive benchmark and check accuracy and bias.

When should I not use time series methods?

When there is no history, the pattern has changed or key events are known that the history does not reflect.