ERP Event Logs: A Canonical Form for KPI Time-Series Extraction and Characterization

arXiv:2610.09903 · cs.DB · Submitted 2026-10-07 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.

Tom: We're looking at the paper: "ERP Event Logs".

Jane: Here's the gist of the paper. ERP event logs provide a canonical form for KPI extraction over relational ERP systems by collapsing schema heterogeneity into a flat structure from which KPIs become reusable operations Pruning redundant reasoning tokens The paper shows that an event log, a canonical (case, activity, timestamp) representation derived from the same tables, collapses this heterogeneity into one flat structure.

Tom: Using standardized event logs built from hundreds of SAP S/4HANA customers under a shared activity ontology, they define three families of process-derived KPIs (volumes, durations, and rates) and extract them uniformly across customer systems for two common end-to-end organizational processes—order-to-cash and procure-to-pay.

Jane: They normalize rates and screen series for minimum coverage to support comparison across customers of very different sizes.

Tom: How the training signal is built The event log records each document that marks the start of an end-to-end process execution as a case, and each processing step as an (activity, timestamp) pair belonging to that case.

Title and authors: Tom: So we’re looking at "ERP Event Logs: A Canonical Form for KPI Time-Series Extraction and Characterization" by Sherri Hadian, Adrian Rebmann, Atacan Korkmaz, Gregor Berg, and Ulf Brackmann. These are the folks who built this framework.

Jane: They are taking something that is spread across hundreds of normalized relational tables in enterprise resource planning systems and trying to find a way to extract key performance indicators from it uniformly.

Lu: The title tells you exactly what they’re doing: creating a canonical form, which is this event log structure with case, activity, and timestamp, to solve the problem of bespoke joins for every metric.

Meng: So why does this matter practically? I’m wondering if this just means we have one more complex layer of data transformation instead of a simpler one.

Lalam: It means they build standardized event logs using a shared activity ontology from hundreds of SAP S/4HANA customers, which allows them to define three families of process derived KPIs: volumes, durations, and rates <ref:2610.09903#pg1,from hundreds of SAP S/4HANA customers>.

Tom: Right, so they are not just looking at raw data points; they are defining these specific kinds of metrics—volumes for how much work flows through a process, durations for how long it takes that work to do something, and rates for how often things go wrong.

Jane: And they extract these uniformly across two major end-to-end processes: order-to-cash and procure-to-pay. That's a big scope because those are huge parts of any business workflow.

Lu: They normalize the rates and screen series for minimum coverage to make sure you can actually compare results across customers who might be very different in size or complexity.

Meng: If you can compare the same type of metric, like lead time, between a small customer’s process and a massive one’s, that gives me some confidence in the comparison part.

Lalam: Exactly. Because order-to-cash and procure-to-pay reduce to the same common structure, their signals become directly comparable without needing to reconcile two totally unrelated join paths.

The paper's summary: Tom: So what’s the actual mechanism they used? They record every document that marks the start of an end-to-end process execution as a case, and every processing step as an activity and a timestamp pair belonging to that case.

Jane: They use this fixed mapping of source records to a canonical activity from a shared vocabulary to resolve the different labels customers might use for the same thing.

Lu: Crucially, they put both kinds of information—the document start and the step details—under one time resolved representation, which recovers series that the state tables had overwritten from history.

Meng: That sounds like they are fixing a known problem where historical data gets messy because newer information overwrites older states.

Lalam: Right, so they are recovering those old series using this single event log structure. It’s not just recording new stuff; it’s reconstructing the process history correctly.

Tom: And then every KPI is then a uniform operation over this one structure, which is why it becomes reusable after that initial mapping effort.

Jane: So, they are essentially creating a standardized language for describing process events so that you can extract these three families of KPIs—volumes, durations, and rates—consistently across different ERP systems.

Lu: They define volume as a count of an activity per period to capture demand and throughput, duration as the timestamp difference between two activities in one case to capture lead times, and rate as a count normalized by a population to capture exceptions.

Meng: I like the distinction between those three families because it gives you different types of operational questions you can ask about the process health.

Lalam: And they are showing how this works for order-to-cash and procure-to-pay, allowing for direct cross process comparison of sales and procurement side signals, including lead times and order arrivals.

The paper's improvements: Tom: Now, the paper points out that the gain isn't just about having more expressive power in relational algebra; it’s about extraction leverage—paying that extraction cost once and reusing KPI operators over one common structure instead of needing a new schema specific join every time.

Jane: That means they are prioritizing the operational efficiency of analysis over just having a theoretically powerful database query language.

Lu: The improvement is that this shared representation yields reusable, comparable, and model ready KPI time series where the ontology aligns semantics and the common schema reuses KPI operators.

Meng: From my side, if it means I don't have to write custom ETL logic for every new dashboard metric against every customer system, that’s a huge win for deployment complexity.

Lalam: It means the ontology aligns the meaning of things across different systems, so when you use a KPI operator once, it works reliably everywhere.

Tom: The paper also characterized the heterogeneity of these processes using simple process descriptors derived directly from this canonical form—like how much work flows through a business process or how long that work takes.

Jane: They are showing how you can characterize the differences between hundreds of customers based on their actual execution patterns, not just what their database schema looks like.

Lu: This allows for direct comparison of sales and procurement side signals, which is something raw ERP would require reconciling two unrelated join paths to do.

Meng: So they are moving the complexity from runtime query time to a one-time mapping phase, which is exactly what we need when shipping models at scale.

Conclusion: Tom: So wrapping up, the core contribution of "ERP Event Logs: A Canonical Form for KPI Time-Series Extraction and Characterization" is framing the event log as this common representation for KPI extraction over relational ERP systems.

Jane: The value is really in that extraction leverage—once mapped, process derived KPIs reduce to reusable operations over one structure, which helps you define three families of KPIs: volumes, durations, and rates.

Lu: They characterized the heterogeneity across hundreds of customers and showed cross process comparability for order-to-cash and procure-to-pay signals.

Meng: I think the most important thing is that this system makes KPI analysis more scalable because it moves the heavy lifting to a one time mapping phase rather than on every query.

Lalam: And they also characterized forecastability patterns across both processes, showing that volume series are the most forecastable family, and forecastability decreases from volumes through durations to rates.

Tom: So the big picture is this: you get a common structure that lets you define uniform KPI families and compare different customer processes easily. This paper really shows how to make KPI extraction more standardized across diverse enterprise systems.

Jane: It sets up a nice foundation for building models that can handle heterogeneity, which is something we all struggle with in real-world data environments.

Lu: The future work would naturally be about testing these forecastability patterns rigorously and seeing if this representation holds up when you introduce even more complex process steps.

Meng: From an engineering standpoint, the challenge for deployment will be making that initial mapping robust enough to handle schema changes in those customer systems without breaking everything.

Lalam: So in summary, "ERP Event Logs: A Canonical Form for KPI Time-Series Extraction and Characterization" gives us a way to treat complex ERP data as a flat structure ready for standardized KPI operations.

SAP SE

cs.DB

Submitted: 2026-10-07

Updated: 2026-11-09

Comments: accepted in D2AI at IEEE ICDM 2026. Version 2 corrects a mislabeled rate family in the forecastability figure (the "receipt rate" row is the PO change rate); clarifies the KPI/rate definitions in Sections I, IV and Table I. Results unchanged

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 77/100

The gist: ERP event logs provide a canonical form for KPI extraction over relational ERP systems by collapsing schema heterogeneity into a flat structure from which KPIs become reusable operations Pruning…

Key concepts

Event Log
A standardized record capturing every start (case) and step (activity) in an end-to-end business process. It collapses the messy, different database structures from various ERP systems into one simple, flat structure suitable for analysis.
KPI Families
Three types of KPIs derived from the event log: volumes (how much work flows), durations (how long work takes), and rates (how often things deviate). These standardized metrics allow for consistent measurement of process health across different business operations.
Extraction Leverage
The main benefit is not new data, but efficiency. By performing the complex extraction once into the common event log structure, subsequent KPI calculations become simple operations on that single structure rather than requiring costly, schema-specific joins for every metric.

Terminology

Summary

ERP event logs provide a canonical form for KPI extraction over relational ERP systems by collapsing schema heterogeneity into a flat structure from which KPIs become reusable operations

Pruning redundant reasoning tokens

The paper shows that an event log, a canonical (case, activity, timestamp) representation derived from the same tables, collapses this heterogeneity into one flat structure. Using standardized event logs built from hundreds of SAP S/4HANA customers under a shared activity ontology, they define three families of process-derived KPIs (volumes, durations, and rates) and extract them uniformly across customer systems for two common end-to-end organizational processes—order-to-cash and procure-to-pay. p. 0 They normalize rates and screen series for minimum coverage to support comparison across customers of very different sizes. p. 0

How the training signal is built

The event log records each document that marks the start of an end-to-end process execution as a case, and each processing step as an (activity, timestamp) pair belonging to that case. Activity labels come from a fixed mapping that resolves each source record to a canonical activity from a shared vocabulary. p. 0 The event log places both kinds under one time-resolved representation, recovering series that the state tables had overwritten from the change history. p. 0

Where the gains come from

The gain is therefore not the expressive power, which relational algebra already has, but extraction leverage: they pay this extraction cost once, and every KPI afterward becomes an operation over one common structure rather than a new schema-specific join. This yields reusable, comparable, and model-ready KPI time series where the ontology aligns semantics and the common schema reuses KPI operators. p. 0

Defining KPI families

The three families of process-derived KPIs correspond to operational questions: how much work flows through a business process, how long that work takes, and how often it deviates. A volume is a count of an activity per period, capturing demand and throughput (order arrivals, completions); a duration is the timestamp difference between two activities in a case, capturing the process lead times; and a rate is a count normalized by a population, capturing the incidence of exceptions such as cancellations or price changes that signal process health. p. 0

Cross-process comparability

Because order-to-cash and procure-to-pay reduce to the same common structure, their signals become directly comparable, a comparison that over raw ERP would require reconciling two unrelated join paths. This shared representation enables direct comparison of sales- and procurement-side signals, including lead times and order arrivals. p. 0

Forecastability patterns

Three patterns hold across both processes regarding forecastability: first, seasonal na¨ıve is the weakest model almost everywhere. Second, volume series are the most forecastable family. p. 4 Third, forecastability decreases from volumes through durations to rates. p. 4

Conclusion

In this paper, they framed the event log as a common representation for KPI extraction over relational ERP systems. Its value is extraction leverage: once mapped, process-derived KPIs reduce to reusable operations over one structure. p. 5 Using this representation, they defined three KPI families and characterized their heterogeneity, cross-process comparability, and forecastability across hundreds of customers. p. 6

References

[1] W. Van Der Aalst et al. p. 6, “Process mining manifesto,” in International conference on business process management. p. 6 p. 6 Springer, 2011, pp. p. 6 p. 6 169–194

[2] W. M. van der Aalst and J. p. 6 Carmona, Process mining handbook. p. 6 Springer, 2022, vol. p. 6 p. 6 p. 6 p. 6 p. 6 448

[3] A Berti et al., “An event data extraction approach from SAP ERP for process mining,” in Process Mining Workshops (ICPM 2021). p. 6 p. 6 p. 6 p. 6 p. 6 Springer, 2021, pp. p. 6 p. 6 p. 6 p. 6 p. 6 255–267 arXiv:2110. p. 6 p. 6 p. 6 p. 6 p. 603467

[4] IEEE, “IEEE standard for eXtensible event stream (XES) for achieving interoperability in event logs and event streams,” IEEE Std 1849-2016, pp. 1–50, 2016

[5] L. J. Tashman, “Out-of-sample tests of forecasting accuracy: An analysis and review,” International Journal of Forecasting, vol. p. 6 16, no. 4, pp. 437–450, 2000

[6] R. J. Hyndman and A. p. 6 B. Koehler, “Another look at measures of forecast accuracy,” International journal of forecasting, vol. p. 6 22, no. 4, pp. 679–688, 2006

[7] R. J. Hyndman et al. p. 6, “A state space framework for automatic forecasting using exponential smoothing methods,” International Journal of forecasting, vol. p. 6 p. 6 p. 6 p. 6 p. 6 18, no. 3, pp. 439–454, 2002

[8] G. E. P. Box et al., Time Series Analysis: Forecasting and Control, 5th ed. p. 6 p. 6 p. 6 p. 6 p. 6 p. 6 Hoboken, NJ, USA: John Wiley & Sons, 2015

[9] A Pankratz, Forecasting with dynamic regression models. John Wiley & Sons, 2012

[10] A F Ansari et al., “Chronos: Learning the language of time series,” Transactions on Machine Learning Research (TMLR), 2024, arXiv:2403. p. 6 p. 6 p. 6 p. 6 p. 6 p. 6 p. 607815

[11] H. L. Lee et al., “Information distortion in a supply chain: The bullwhip effect,” Management Science, vol. p. 6 p. 6 p. 6 p. 6 p. 6 p. 6 43, no. 4, pp. 546–558, 1997

[12] A F Ansari et al., “Chronos: Learning the language of time series,” Transactions on Machine Learning Research (TMLR), 2024, arXiv:2403.07815

[13] R J Hyndman and A B Koehler, “Another look at measures of forecast accuracy,” International journal of forecasting, vol. 22, no. 4, pp. 679–688, 2006

[14] G E P Box et al., Time Series Analysis: Forecasting and Control, 5th ed. Hoboken, NJ, USA: John Wiley & Sons, 2015

[15] A Pankratz, Forecasting with dynamic regression models. John Wiley & Sons, 2012

[16] A F Ansari et al., “Chronos: Learning the language of time series,” Transactions on Machine Learning Research (TMLR), 2024, arXiv:2403.07815

[17] H L Lee et al., “Information distortion in a supply chain: The bullwhip effect,” Management Science, vol. 43, no. 4, pp. 546–558, 1997

[18] W Van Der Aalst et al., “Process mining manifesto,” in International conference on business process management. Springer, 2011, pp. 169–194

[19] W M van der Aalst and J Carmona, Process mining handbook. Springer, 2022, vol. 448

[20] A Berti et al., “An event data extraction approach from SAP ERP for process mining,” in Process Mining Workshops (ICPM 2021). Springer, 2021, pp. 255–267 arXiv:2110.03467

[4] IEEE, “IEEE standard for eXtensible event stream (XES) for achieving interoperability in event logs and event streams,” IEEE Std 1849-2016, pp. 1–50, 2016

[5] L J Tashman, “Out-of-sample tests of forecasting accuracy: An analysis and review,” International Journal of Forecasting, vol. 16, no. 4, pp. 437–450, 2000

[6] R J Hyndman and A B Koehler, “Another look at measures of forecast accuracy,” International journal of forecasting, vol. 22, no. 4, pp. 679–688, 2006

[7] R J Hyndman et al., “A state space framework for automatic forecasting using exponential smoothing methods,” International Journal of forecasting, vol. 18, no. 3, pp. 439–454, 2002

[8] G E P Box et al., Time Series Analysis: Forecasting and Control, 5th ed. Hoboken, NJ, USA: John Wiley & Sons, 2015

[9] A Pankratz, Forecasting with dynamic regression models. John Wiley & Sons, 2012

[10] A F Ansari et al., “Chronos: Learning the language of time series,” Transactions on Machine Learning Research (TMLR), 2024, arXiv:2403.07815

[11] H L Lee et al., “Information distortion in a supply chain: The bullwhip effect,” Management Science, vol. 43, no. 4, pp. 546–558, 1997

[2] W M van der Aalst and J Carmona, Process mining handbook. Springer, 2022, vol. 448

[3] A Berti et al., “An event data extraction approach from SAP ERP for process mining,” in Process Mining Workshops (ICPM 2021). Springer, 2021, pp. 255–267 arXiv:2110.03467

[4] IEEE, “IEEE standard for eXtensible event stream (XES) for achieving interoperability in event logs and event streams,” IEEE Std 1849-2016, pp. 1–50, 2016

[5] L J Tashman, “Out-of-sample tests of forecasting accuracy: An analysis and review,” International Journal of Forecasting, vol. 16, no. 4, pp. 437–450, 2000

[6] R J Hyndman and A B Koehler, “Another look at measures of forecast accuracy,” International journal of forecasting, vol. 22, no. 4, pp. 679–688, 2006

[7] R J Hyndman et al., “A state space framework for automatic forecasting using exponential smoothing methods,” International Journal of forecasting, vol. 18, no. 3, pp. 439–454, 2002

[8] G E P Box et al., Time Series Analysis: Forecasting and Control, 5th ed. Hoboken, NJ, USA: John Wiley & Sons, 2015

[9] A Pankratz, Forecasting with dynamic regression models. John Wiley & Sons, 2012

[10] A F Ansari et al., “Chronos: Learning the language of time series,” Transactions on Machine Learning Research (TMLR), 2024, arXiv:2403.07815

[11] H L Lee et al., “Information distortion in a supply chain: The bullwhip effect,” Management Science, vol. 43, no. 4, pp. 546–558, 1997

[2] W M van der Aalst and J Carmona, Process mining handbook. Springer, 2022, vol. 448

[3] A Berti et al., “An event data extraction approach from SAP ERP for process mining,” in Process Mining Workshops (ICPM 2021). Springer, 2021, pp. 255–267 arXiv:2110.03467

[4] IEEE, “IEEE standard for eXtensible event stream (XES) for achieving interoperability in event logs and event streams,” IEEE Std 1849-2016, pp. 1–50, 2016

[5] L J Tashman, “Out-of-sample tests of forecasting accuracy: An analysis and review,” International Journal of Forecasting, vol. 16, no. 4, pp. 437–450, 2000

[6] R J Hyndman and A B Koehler, “Another look at measures of forecast accuracy,” International journal of forecasting, vol. 22, no. 4, pp. 679–688, 2006

[7] R J Hyndman et al., “A state space framework for automatic forecasting using exponential smoothing methods,” International Journal of forecasting, vol. 18, no. 3, pp. 439–454, 2002

[8] G E P Box et al., Time Series Analysis: Forecasting and Control, 5th ed. Hoboken, NJ, USA: John Wiley & Sons, 2015

[9] A Pankratz, Forecasting with dynamic regression models. John Wiley & Sons, 2012

[10] A F Ansari et al., “Chronos: Learning the language of time series,” Transactions on Machine Learning Research (TMLR), 2024, arXiv:2403.07815

[11] H L Lee et al., “Information distortion in a supply chain: The bullwhip effect,” Management Science, vol. 43, no. 4, pp. 546–558, 1997

[2] W M van der Aalst and J Carmona, Process mining handbook. Springer, 2022, vol. 448

[3] A Berti et al., “An event data extraction approach from SAP ERP for process mining,” in Process Mining Workshops (ICPM 2021). Springer, 2021, pp. 255–267 arXiv:2110.03467

[4] IEEE, “IEEE standard for eXtensible event stream (XES) for achieving interoperability in event logs and event streams,” IEEE Std 1849-2016, pp. 1–50, 2016

[5] L J Tashman, “Out-of-sample tests of forecasting accuracy: An analysis and review,” International Journal of Forecasting, vol. 16, no. 4, pp. 437–450, 2000

[6] R J Hyndman and A B Koehler, “Another look at measures of forecast accuracy,” International journal of forecasting, vol. 22, no. 4, pp. 679–688, 2006

[7] R J Hyndman et al., “A state space framework for automatic forecasting using exponential smoothing methods,” International Journal of forecasting, vol. 18, no. 3, pp. 439–454, 2002

[8] G E P Box et al., Time Series Analysis: Forecasting and Control, 5th ed. Hoboken, NJ, USA: John Wiley & Sons, 2015

[9] A Pankratz, Forecasting with dynamic regression models. John Wiley & Sons, 2012

[10] A F Ansari et al., “Chronos: Learning the language of time series,” Transactions on Machine Learning Research (TMLR), 2024, arXiv:2403.07815

[11] H L Lee et al., “Information distortion in a supply chain: The bullwhip effect,” Management Science, vol. 43, no.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Canonical Representation for KPI Extraction: Implement an event log structure (case, activity, timestamp) derived from normalized relational ERP tables to collapse heterogeneity into one flat structure. This enables extraction leverage by paying the extraction cost once and reusing KPI operators over this common structure instead of requiring a bespoke join for every metric.

  2. Uniform KPI Definition: Define three families of process-derived KPIs—volumes, durations, and rates—as uniform operations over the event log to ensure they are comparable across tenants and across order-to-cash and procure-to-pay. This standardization allows for consistent analysis regardless of the underlying source schema.

  3. Heterogeneity Characterization: Use simple process descriptors derived directly from the canonical form, such as the number of cases (process volume), the number of distinct activities (process complexity), and the median completed-case duration, to characterize process heterogeneity across hundreds of customers and industries. This allows for direct comparison of how different customer processes execute an end-to-end workflow.

  4. Cross-Process Signal Comparison: Enable direct cross-process comparison by leveraging the shared structure, allowing for direct cross-process comparison of sales and procurement side signals, including lead times and order arrivals. This contrasts with raw ERP which requires reconciling two unrelated join paths to compare these metrics.

  5. Forecastability Analysis: Assess forecastability uniformly across KPI families, processes, and customers using models like Chronos against classical baselines (ARIMA, ETS). This system can determine which KPI families are forecastable, providing a recurring ordering from volumes through durations to rates across both processes and customer bases.

Sources

Related papers