How Exito Transformed Media Spend Forecasting with TabPFN

Background

exito GmbH & Co. KG, one of Germany’s leading online marketing agencies, is leveraging TabPFN to improve the accuracy and efficiency of its media spend forecasting. Exito's data science team develops the databases and analytical tools that power their account managers' campaign planning and execution.

They use first-party data from CSV files or, for example, customer ERP systems and marketing data to optimize spend, campaign performance, and ROI. Forecasting plays a central role in this process: accurate daily and monthly spend predictions that would allow Exito to anticipate customer budgets, adjust campaigns dynamically, and project internal revenues - since their business model depends directly on total customer ad spend.

By integrating TabPFN, Exito has built a more efficient forecasting pipeline and significantly improved daily forecast accuracy.

The Challenge

Forecasting advertising spend presents multiple layers of complexity. Each customer’s behavior differs, and spend patterns fluctuate due to seasonality, media type, and platform dynamics. Major events like Black Friday or Cyber Monday can also skew patterns, making it difficult to maintain accuracy across customers. Most importantly, specific discount campaigns, promotions or budget-limitations on specific dates cannot be anticipated by any model.

Exito’s previous forecasting setup relied on naive methods broadly used in marketing like extrapolating costs linearly or taking into account last year’s spend. They also did some experiments with Prophet, XGBoost and neural networks. While Prophet offered interpretability, its accuracy fell short - especially when forecasting multiple customers simultaneously. The model also required significant tuning to achieve acceptable results.

Key bottlenecks:

• Manual model tuning per customer dataset
• Inconsistent accuracy & inability to generalize across customers
• Inability to account for customer specific discounts or promotions

“We had up to three years of hourly spend data per customer, but even with consistent data, the old models often missed monthly totals and were in general worse than TabPFN (e.g. in MAPE) or reacted too slowly to short-term shifts”, says Hendrik Dohme, data scientist at Exito.

Introducing TabPFN

To address these limitations, Exito began experimenting with TabPFN, a transformer-based foundation model for tabular data developed by Prior Labs. TabPFN was tested on the same three years of daily media spend data and immediately stood out - outperforming Prophet, neural networks, and XGBoost in nearly all cross-validation scenarios, despite using less granular data.

“Even without hourly data, TabPFN gave us significantly better results on daily forecasts. The quality of short-term predictions was much higher”, says Dohme.

From Manual Forecasting to Automated Insight

In the new setup, Exito uses TabPFN via API to predict daily media spend for the remaining days of the month for each customer, updating forecasts every 2 hours.

• Exogenous variables are included such as the available hourly spend of the current day which was not possible with previous models. These change with every forecast.
• Each prediction task runs in a few seconds.
• 19 parallel forecasts are run from 8am to 6pm every two hours.
• Predictions can be aggregated to estimate total monthly spend - the key KPI for both campaign planning and internal revenue projection.

TabPFN handles new data seamlessly - requiring no special preprocessing for categorical data, no hyperparameter tuning, and minimal maintenance.

Impact on the Workflow

With other forecasting models like Prophet, each new customer or campaign type required manual model selection and tuning. With TabPFN, Exito fully automated its forecasting workflow.

• Deployment: Exito integrated TabPFN through the hosted API without any GPU infrastructure, fitting directly into their existing sci-kit learn pipeline.
• Scalability: Each forecast runs in under a minute, enabling updates multiple times per day across all customers.
• Efficiency: Forecasting tasks are now generalized across customers, cutting model development time from hours to minutes.
• Automation: TabPFN requires no retraining or manual tuning. New data is automatically incorporated as context, allowing the model to adapt continuously.

Business Value

Better forecasts mean better campaign decisions - and more accurate internal revenue projections.

• Exito can anticipate when spend will spike or drop, adjusting campaigns proactively to maximize ROI.
• The agency can project its own revenue with greater confidence, since it’s tied directly to customer spend thresholds.
• Forecasting insights empower account managers to optimize budgets dynamically - for example, reducing spend on rainy days when conversion drops for certain product categories.

“With TabPFN, we’re able to move faster, predict better, and rely less on manual modeling. It’s a step toward true automation in marketing forecasting”, says Dohme.

Conclusion

Exito’s integration of TabPFN demonstrates how a tabular foundation model can transform digital marketing analytics by combining speed, automation, and accuracy.

By replacing manual tuning with an API-based workflow, Exito achieved faster iteration cycles, more reliable forecasts, and greater visibility into customer performance. The success of this implementation underscores the growing potential of TabPFN to enhance predictive modeling in data-rich but noisy domains such as marketing, finance, and operations.

Ready to transform your forecasting workflow? Learn more about Prior Labs or discover Exito's online marketing solutions.

‍