Methodology and design principles A Clarasight white paper

Forecasting Business Travel Spend and Emissions

A forecasting methodology for Finance teams to update travel spend and emissions forecasts as revenue plans, actual spend, policies, and costs change.

What's inside

A monthly model that learns repeating patterns from historical data and refreshes as new actuals arrive.
The spend-to-revenue ratio used to fold in revenue projections without adding model complexity.
Three scenario types for comparing several futures side by side in one sitting.

Read the methodology

Methodology

1. Overview

Forecasting corporate travel spend and emissions is a deceptively complex problem. Travel data is seasonal, noisy, and sensitive to both business conditions and one-off events. A useful forecast has to stay stable and interpretable over a multi-year horizon, update gracefully as new actuals come in, and support scenario planning around policy levers and macroeconomic assumptions.

This paper outlines the core methodology Clarasight uses to build long-range forecasts of business travel spend and emissions. The approach is built around a deliberate set of modeling choices, each made to address a specific challenge in travel data. The result is a forecasting system that is accurate, transparent, and flexible enough to support both annual planning and mid-year monitoring.

Much of today's forecasting conversation centers on AI. Clarasight's approach draws from a broader statistical toolkit. The method in this paper is deliberately conventional: a proven statistical workhorse with visible assumptions and behavior that can be tested against actuals. For long-range planning, that predictability is a strength.

2. Why (S)ARIMA?

The foundation of the approach is the (S)ARIMA — (Seasonal) AutoRegressive Integrated Moving Average — model family. These models allow for a forecast that accounts for observed annual seasonality while keeping scenario modeling responsive enough to provide clear visibility for decision-making.

Corporate travel spend is seasonal in a way that is very company-specific. A consulting organization might see relatively consistent spend outside the holidays, while a company with conference-dependent travel sees high spend at certain times of year and little to none in others. Even within one company, billable and non-billable travel can move differently, as can spend across modalities like air and hotel. SARIMA models handle seasonality natively through their seasonal components, learning these repeating patterns from historical data rather than requiring manual adjustment. Because the model is computationally simple, generating multiple forecasts across different slices of the business — each with its own temporal pattern — is a trivial exercise.

At the same time, a long-range forecast needs to fold in year-to-date actuals to update its rest-of-year projection and pick up any shift in seasonality. SARIMA is designed for exactly this: it trains on a monthly time series and produces a monthly forecast, so a mid-year refresh simply means re-running the model on a training set that includes the recent actuals. The model structure — defined by up to six parameters — can even evolve as more data arrives: seasonal patterns get reinforced or dampened, and policy shifts soak in.

Figures 1 and 2 show what that refresh looks like when the latest five months of actual spend run 20% above the prior series: the monthly model is re-fitted on the elevated data, and the annual view shows how that change carries into projected totals.

Monthly time-series chart of illustrative travel spend. The latest five months of actuals run 20% above the baseline series, and a re-fitted forecast diverges from the baseline forecast. Illustrative data. — FIGURE 1. Year-to-date refresh, monthly view. The latest five months of actual spend are 20% above the baseline series, and the model is re-fitted on the elevated data. Illustrative data

Annual view comparing baseline and re-fitted forecasts of illustrative travel spend, showing how elevated year-to-date actuals carry into annual projected totals. Illustrative data. — FIGURE 2. Year-to-date refresh, annual view. Baseline and re-fitted forecasts show how the elevated actuals carry into annual projected spend. Illustrative data

Inputs

3. Prediction variables: business inputs and market forces

Prediction variables beyond time itself fall into two classes: business inputs internal to the company (endogenous variables) and market forces external to it (exogenous variables). Clarasight's forecasting tools handle the two differently.

Business inputs

Businesses often have robust internal projections for endogenous variables like headcount and revenue, driven by market research and product roadmaps. Clarasight can incorporate these directly and take a company's own projections at their word.

At Clarasight, we made the design choice to forecast the ratio of monthly spend to annual revenue (or a similar scaling variable), rather than raw spend figures directly. This decision solves several problems at once.

First, we have observed across several companies that raw travel spend tends to grow over time as a company grows its revenue. Comparable public data on corporate travel cost is scarce. Figure 3 therefore uses disclosed business-travel emissions as a more consistently available proxy for travel activity, paired with reported revenue. Each series shows the change from its 2022 value.

FIGURE 3. Company revenue and business-travel emissions change versus 2022 across Microsoft, Accenture, Amazon, and Meta. Public company reports; fiscal years where applicable.

Modeling raw spend or emissions would require the model to learn and extrapolate an underlying revenue (an annual or quarterly phenomenon) to spend/emissions trend (a monthly phenomenon), which is difficult to do reliably over a multi-year horizon. By dividing the variable of interest by revenue, we remove the revenue-based component of the trend, meaning that the ratio is more likely to be stationary (roughly constant in expectation over time), which is exactly the structural property ARIMA models are designed to exploit.

Second, this structure makes it straightforward to incorporate revenue projections into the forecast. Once the model produces a predicted spend/revenue ratio for each future month, the dollar forecast is simply recovered by multiplying by projected annual revenue. This means that changing the revenue growth assumption (say, updating from 6% to 3%, or considering non-constant growth due to M&A activity) requires no model retraining. We can model arbitrarily complex business trajectories over time without increasing model complexity — particularly valuable for scenario planning.

Finally, it produces more stable long-range forecasts. A model that learns "spend is X% of revenue in month Y of the fiscal year" generalizes more robustly to future years than one that tries to extrapolate an absolute dollar trend.

External market forces

Exogenous variables like oil prices and inflation can influence travel cost, but the relationship is often unclear when regressing on past data. To include an exogenous variable as a regression input in a useful, meaningful way, three requirements will ideally hold:

The external variable itself can be reasonably predicted over the forecast period. For inflation, oil prices, or airfares, that is a hard problem on its own.
The external variable is not multicollinear with company-specific internal variables already in the model, such as revenue. Including several variables that encode the same underlying phenomenon adds nothing to a forecast. A short thought exercise shows how this condition can rule out many candidates: oil prices affect every sector of the economy and directly influence the Consumer Price Index; CPI and its growth drive the salary and overhead a business carries, which in turn shape the revenue it needs to hold its margins. Add oil as a prediction variable on top of revenue, and the model often encodes the same underlying force twice.
The company's past response to the variable resembles how it will respond in future. To model travel costs falling as oil prices rise, a company must have already shown that responsiveness historically. In reality, travel demand is often inelastic except under large, historically unseen shocks.

Because these conditions are difficult to satisfy, our research indicates inferior backtesting performance and more model instability when these variables are included. The approach can still account for their effects — but it does so by modeling how the business intends to respond, rather than trying to predict the shock itself. That response might be a travel freeze for a set number of months, a cap on spend, or no change at all. In practice, teams can compare the refreshed forecast with actual spend and budget throughout the year.

4. Sensitivity analysis: planning for uncertainty

Because the SARIMA framework is inexpensive to re-run, it is well suited to structured sensitivity analysis. Clarasight's forecasting supports three kinds.

Revenue sensitivity shows the impact of varying the revenue growth assumptions across a range of scenarios. Because the cost-to-revenue ratio remains identical across all growth scenarios, only the scale of the projections changes — easy to communicate a range of outcomes to business stakeholders without separate model re-runs.

Policy-change sensitivity models the impact of a structural initiative, such as shifting a class of air bookings to a lower fare category. It works by adjusting the historical training series — modifying observed data points and re-fitting the model — to quantify a counterfactual in which the policy had always been in effect. The result is a structurally similar forecast with a potentially different seasonality and cost-to-revenue ratio, useful for estimating a policy's impact in future years.

Figures 4 and 5 show that counterfactual in practice: an initiative removes roughly 10% of modeled travel spend, and the re-fitted model carries the effect from the monthly pattern into annual totals.

Monthly time-series chart of illustrative travel spend under a policy change: retained travel, travel eliminated by the initiative, and baseline versus post-initiative forecasts under a roughly 10% reduction. Illustrative data. — FIGURE 4. Policy-change sensitivity, monthly view. Retained travel, travel eliminated by the initiative, and baseline versus post-initiative forecasts under a roughly 10% reduction. Illustrative data

Annual view of the policy-change scenario showing historical retained and eliminated travel alongside baseline and post-initiative annual forecasts of illustrative travel spend. Illustrative data. — FIGURE 5. Policy-change sensitivity, annual view. Historical retained and eliminated travel alongside baseline and post-initiative annual forecasts. Illustrative data

Cost-shock sensitivity introduces an artificial change in one or more past or future data points, simulating a one-time event — a freeze, a policy shift, an operational shock — that drives spend anomalously low in a single month. As with a policy change, a counterfactual model is fit and used to project the shock's effect forward.

The same counterfactual method can also combine events rather than test them one at a time. Figures 6 and 7 follow a five-month 20% spend spike followed by an enforced 15% reduction through year-end, then compare the original baseline, the post-spike path, and the post-cap re-fit month by month and annually.

Monthly time-series chart of an illustrative compound scenario: a 20% spend spike from January through May 2026 followed by a 15% cap from June through December, with business-as-usual, post-spike, and post-cap re-fitted forecasts. Illustrative data. — FIGURE 6. Compound sensitivity, monthly view. A 20% spend spike from January through May 2026 is followed by a 15% cap from June through December, with business-as-usual, post-spike, and post-cap re-fitted forecasts. Illustrative data

Annual view of the compound 2026 scenario, rolling the three forecast paths up to annual totals for illustrative travel spend. Illustrative data. — FIGURE 7. Compound sensitivity, annual view. The mixed 2026 scenario and the three forecast paths rolled up to annual totals. Illustrative data

Reference

Glossary

Key terms

SARIMA

Seasonal AutoRegressive Integrated Moving Average. A standard time-series model that learns repeating seasonal patterns from monthly data.

Seasonality

The repeating within-year pattern in a company's travel: busy months and quiet months, specific to each company.

Stationarity

A series whose behavior stays roughly constant over time, with no persistent upward or downward drift. Time-series models forecast stationary series more reliably — which is why this approach forecasts a ratio rather than raw dollars.

Next step

See Clarasight in action.

Explore how Clarasight connects travel data, planning assumptions, and forecasts in one workflow.

Request a demo