Skip to main content
Disciplines
Automotive Business IT + Informatics Construction + Real Estate Electrical Engineering + Electronics Energy + Sustainability Insurance + Risk Finance + Banking Management + Leadership Marketing + Sales Mechanical Engineering + Materials
Events
DE
EN
Books
Journals
Topic Page
Marketing
Start single access now
Access for companies
EXTENDED SEARCH
Log in
Top
Review of Accounting Studies
Published in: Review of Accounting Studies 3/2020

07-07-2020

Machine learning improves accounting estimates: evidence from insurance payments

Authors: Kexing Ding, Baruch Lev, Xuan Peng, Ting Sun, Miklos A. Vasarhelyi

Published in: Review of Accounting Studies | Issue 3/2020

Log in

Activate our intelligent search to find suitable subject content or patents.

search-config
loading …

Abstract

Managerial estimates are ubiquitous in accounting: most balance sheet and income statement items are based on estimates; some, such as the pension and employee stock options expenses, derive from multiple estimates. These estimates are affected by objective estimation errors as well as by managerial manipulation, thereby harming the reliability and relevance of financial reports. We show that machine learning can substantially improve managerial estimates. Specifically, using insurance companies’ data on loss reserves (future customer claims) estimates and realizations, we document that the loss estimates generated by machine learning were superior to actual managerial estimates reported in financial statements in four out of five insurance lines examined. Our evidence suggests that machine learning techniques can be highly useful to managers and auditors in improving accounting estimates, thereby enhancing the usefulness of financial information to investors.
previous article Stock compensation expense, cash flows, and inflated valuations
next article Machine learning improves accounting: discussion, implementation and research opportunities

Dont have a licence yet? Then find out more about our products and how to get one now:

Springer Professional "Wirtschaft+Technik"

Online-Abonnement

Mit Springer Professional "Wirtschaft+Technik" erhalten Sie Zugriff auf:

  • über 102.000 Bücher
  • über 537 Zeitschriften

aus folgenden Fachgebieten:

  • Automobil + Motoren
  • Bauwesen + Immobilien
  • Business IT + Informatik
  • Elektrotechnik + Elektronik
  • Energie + Nachhaltigkeit
  • Finance + Banking
  • Management + Führung
  • Marketing + Vertrieb
  • Maschinenbau + Werkstoffe
  • Versicherung + Risiko

Jetzt Wissensvorsprung sichern!

inform now

Springer Professional "Wirtschaft"

Online-Abonnement

Mit Springer Professional "Wirtschaft" erhalten Sie Zugriff auf:

  • über 67.000 Bücher
  • über 340 Zeitschriften

aus folgenden Fachgebieten:

  • Bauwesen + Immobilien
  • Business IT + Informatik
  • Finance + Banking
  • Management + Führung
  • Marketing + Vertrieb
  • Versicherung + Risiko




Jetzt Wissensvorsprung sichern!

inform now
Appendix
Available only for authorised users
Footnotes
1
The literature has identified various managerial incentives may impact insurers’ reserve manipulation, including tax deferral (Grace 1990; Petroni 1992; Nelson 2000), income smoothing (Anderson 1973; Smith 1980; Weiss 1985; Beaver, McNichols, and Nelson 2003), solvency and regulatory concerns (Forbes 1970; Petroni 1992; Nelson 2000; Gaver and Paterson 2004; Hoyt and McCullough 2010), and executive compensation incentives (Browne et al. 2009; Eckles and Halek 2010;). We provide more detailed discussions on managerial incentives in Section 7.
 
2
Hyper-parameter is a parameter whose value is set before the learning process begins. Setting up the value of a hyper-parameter controls the process of defining the model.
 
3
It is generally believed that adding more trees in random forest will not introduce overfitting because each individual tree has limited depth (Breiman 2001). The performance of the random forest tends to stay stable at a certain value after a certain number of trees.
 
4
Activation functions are mathematical equations that determine the output of a neural network. The function is attached to each node in the neural network to determine whether the node should be activated. A node will be activated when the output exceeds a certain threshold based on the activation function. When the node is activated, the output in the node will be transmitted to the next layer in the neural network.
 
5
For example, we first used data from 1996 to 2005 to develop machine learning models and applied the best-performing model to generate predictions for the year 2006. We have also tried dividing the sample in other ways, such as using the years 1996–2007 as the training sample and the final year 2008 as the testing set or using the entire data as the sample for cross-validation without holdout. The results were mostly consistent with our reported findings. We report the design of three holdout periods because it provides a more robust evaluation of model performance, given the size of our sample, while accounting for the inherent time-series of the data.
 
6
State laws and insurance regulations require that insurance companies operating in the United States and its territories prepare statutory financial statements in accordance with SAP. SAP is designed to assist state insurance departments in regulating insurance companies’ solvency.
 
7
SAP in general do not require firms to discount losses so that the estimates and actual payments are comparable. Some firms may choose to implicitly discount the future payments to reduce the reported reserve, especially for long-tail lines. However, as the inherent discount rate is not disclosed and discounting is not a standard procedure applied to all firms or all business lines, we calculated our dependent variable undiscounted. We also acknowledge the possibility that loss claims may take longer than 10 years to settle, in which case the cumulative payments in yeart + 9 do not fully capture the actual losses. To alleviate the concern, we first discuss the payment patterns of each business line studied in the next section. Second, we used the manager estimates in yeart + 9 as our proxy for actual losses instead, and the main inferences remained unchanged.
 
8
Insurance Information Institute, The Insurance Fact Book 2017.
 
9
The example was drawn from the period 1996–2007, with 1996–2006 as the cross-validation set and 2007 the holdout set.
 
10
The data from S&P Global Market Intelligence platform is presented in thousands of US dollars. Therefore results are expressed in thousands.
 
11
The cross-validation test results of other machine learning models are provided in Appendix Table 10.
 
12
We thank an anonymous reviewer for suggesting this test. For brevity, we only report the important variables for models trained during the sample period 1996–2007 because the influential variables appear to be similar across different samples for each line.
 
13
For completeness, we also show holdout prediction results for other models in Appendix Table 12. However, the results in appendix Table 12 are used for model selection.
 
14
The model performance downturn in 2008 for line No. 4 might be caused by the economic turbulence during the financial crisis years, 2007–2009, which interrupted the patterns of insurance loss payments. The National Bureau of Economic Research (NBER) marks December 2007–June 2009 as a peak recession period.
 
15
We thank an anonymous reviewer for suggesting this test.
 
16
We thank an anonymous reviewer for suggesting the analysis. We acknowledge that this estimate is not equivalent to the loss reserves in insurers’ financial statement, because the latter may include reserves for business lines other than the five examined as well as for prior years’ losses. However, the other business lines are relatively minor, without sufficient data to support the model development, and estimating the loss reserves for prior years will impose much stricter requirements on our sample period, resulting in a sample size too small to conduct analyses. Therefore we added the five business lines to approximate the true reserve level.
 
18
We note that, while Net Income and Loss Reserve both involve managers’ estimates in calculation, the error components cancel out when we added the two items together so that the total approximates the taxable income before reserves are determined.
 
19
As net income is a primary component of surplus changes, regulators are implicitly encouraging smoother earnings (Grace 1990).
 
20
We have re-estimated the regressions excluding the workers’ compensation line and using linear regression algorithm predictions for the homeowner/farmowner line, and the empirical inferences remain unchanged.
 
Literature
A. M. Best Company. (1994). Best’s aggregates and averages: Property-casualty edition. Oldwick: A. M. Best Company.
Anderson, D. R. (1971). Effects of under and overevaluations in loss reserves. Journal of Risk and Insurance, 585–600.
Anderson, D. R. (1973). Effects of loss reserve evaluation upon policyholders’ surplus. Madison, Wisconsin: Bureau of Business Research and Service, University of Wisconsin, monograph, 6.
Bao, Y., Ke, B., Li, B., Yu, Y. J., & Zhang, J. (2020). Detecting accounting fraud in publicly traded US firms using a machine learning approach. Journal of Accounting Research, 58(1), 199–235.CrossRef
Barboza, F., Kimura, H., & Altman, E. (2017). Machine learning models and bankruptcy prediction. Expert Systems with Applications, 83, 405–417.CrossRef
Beaver, W. H., & McNichols, M. F. (1998). The characteristics and valuation of loss reserves of property casualty insurers. Review of Accounting Studies, 3(1–2), 73–95.CrossRef
Beaver, W. H., McNichols, M. F., & Nelson, K. K. (2003). Management of the loss reserve accrual and the distribution of earnings in the property-casualty insurance industry. Journal of Accounting and Economics, 35(3), 347–376.CrossRef
Bertomeu, J., Cheynel, E., Floyd, E., & Pan, W. (2020). Using machine learning to detect misstatements. Review of Accounting Studies, forthcoming.
Bierens, H. J., & Bradford, D. F. (2005). Are property-casualty insurance reserves biased? A Non-Standard Random Effects Panel Data Analysis 1.
Bishop, C. M. (2006). Pattern recognition and machine learning: Springer.
Breiman, L. (2002). Using models to infer mechanisms. IMS Wald Lecture, 2, 59–71.
Browne, M. J., Ma, Y.-L., & Wang, P. (2009). Stock-based executive compensation and reserve errors in the property and casualty insurance industry. Journal of Insurance Regulation, 27(4).
Eckles, D. L., & Halek, M. (2010). Insurer reserve error and executive compensation. Journal of Risk and Insurance, 77(2), 329–346.CrossRef
Fischer, P. E., & Verrecchia, R. E. (2000). Reporting bias. The Accounting Review, 75(2), 229–245.CrossRef
Forbes, S. W. (1970). Loss reserving performance within the regulatory framework. Journal of Risk and Insurance, 527–538.
Friedman, J. H. (2001). Greedy function approximation: A gradient boosting machine. Annals of Statistics, 1189–1232.
Gaver, J. J., & Paterson, J. S. (2004). Do insurers manipulate loss reserves to mask solvency problems? Journal of Accounting and Economics, 37(3), 393–416.CrossRef
Grace, E. V. (1990). Property-liability insurer reserve errors: A theoretical and empirical analysis. Journal of Risk and Insurance, 28–46.
Grace, M. F., & Leverty, J. T. (2012). Property–liability insurer reserve error: Motive, manipulation, or mistake. Journal of Risk and Insurance, 79(2), 351–380.CrossRef
Guttman, I., & Marinovic, I. (2018). Debt contracts in the presence of performance manipulation. Review of Accounting Studies, 23(3), 1005–1041.CrossRef
Hoyt, R. E., & McCullough, K. A. (2010). Managerial discretion and the impact of risk-based capital requirements on property-liability insurer reserving practices. Journal of Insurance Regulation, 29(2).
Lambert, R. A. (1984). Income smoothing as rational equilibrium behavior. Accounting Review, 604–618.
Mason, L., Baxter, J., Bartlett, P. L., & Frean, M. R. (2000) Boosting algorithms as gradient descent. In Advances in neural information processing systems, (pp. 512–518).
Nelson, K. K. (2000). Rate regulation, competition, and loss reserve discounting by property-casualty insurers. The Accounting Review, 75(1), 115–138.CrossRef
Perols, J. (2011). Financial statement fraud detection: An analysis of statistical and machine learning algorithms. Auditing: A Journal of Practice & Theory, 30(2), 19–50.CrossRef
Perols, J. L., Bowen, R. M., Zimmermann, C., & Samba, B. (2017). Finding needles in a haystack: Using data analytics to improve fraud prediction. The Accounting Review, 92(2), 221–245.CrossRef
Petroni, K., & Beasley, M. (1996). Errors in accounting estimates and their relation to audit firm type. Journal of Accounting Research, 34(1), 151–171.CrossRef
Petroni, K. R. (1992). Optimistic reporting in the property-casualty insurance industry. Journal of Accounting and Economics, 15(4), 485–508.CrossRef
Samuels, D., Taylor, D. J., & Verrecchia, R. E. (2018). Financial misreporting: Hiding in the shadows or in plain sight? Available at SSRN, 3157222.
Schapire, R. E. (1990). The strength of weak learnability. Machine Learning, 5(2), 197–227.
Smith, B. D. (1980). An analysis of auto liability loss reserves and underwriting results. Journal of Risk and Insurance, 305–320.
Sun, T., & Vasarhelyi, M. A. (2017). Deep learning and the future of auditing: How an evolving technology could transform analysis and improve judgment. CPA Journal, 87(6).
Trueman, B., & Titman, S. (1988). An explanation for accounting income smoothing. Journal of Accounting Research, 127–139.
Weiss, M. (1985). A multivariate analysis of loss reserving estimates in property-liability insurers. Journal of Risk and Insurance, 199–221.
Witten, I. H., Frank, E., & Hall, M. A. (2011). Data mining: Practical machine learning tools and techniques. 3rd Edition, Burlington: Morgan Kaufmann Publishers.
Zhang, C., & Browne, M. J. (2013) Loss reserve errors, income smoothing and firm risk of property and casualty insurance companies. In Annual Meeting of the American Risk and Insurance Association, Working Pa, (pp. 1–55).
Metadata
Title
Machine learning improves accounting estimates: evidence from insurance payments
Authors
Kexing Ding
Baruch Lev
Xuan Peng
Ting Sun
Miklos A. Vasarhelyi
Publication date
07-07-2020
Publisher
Springer US
Published in
Review of Accounting Studies / Issue 3/2020
Print ISSN: 1380-6653
Electronic ISSN: 1573-7136
DOI
https://doi.org/10.1007/s11142-020-09546-9

Other articles of this Issue 3/2020

Review of Accounting Studies 3/2020 Go to the issue

Invited Submission

Machine learning improves accounting: discussion, implementation and research opportunities

OriginalPaper

The effects of MiFID II on sell-side analysts, buy-side analysts, and firms

OriginalPaper

Stock-based compensation, financial analysts, and equity overvaluation

OriginalPaper

Stock compensation expense, cash flows, and inflated valuations

OriginalPaper

Measuring disclosure using 8-K filings

OriginalPaper

High-frequency traders and price informativeness during earnings announcements