**5. New challenges in financial modeling**

Despite the numerous opportunities and advantages that machine learning and artificial intelligence-based applications are likely to provide to the financial sector, there will be some initial challenges and barriers too. This section highlights some of the challenges as follows.

**Data challenges:** While the availability of data in finance is quite plenty, the time series data in finance (e.g., stock prices) are quite small in size for data-hungry machine learning and deep learning models. Models built on limited time series data are naturally less trained and improperly designed. The result is a sub-optimal performance of the models. Another problem in finance is that financial data cannot be synthesized unlike images in the fields of computer vision and image processing. Since finance data cannot be synthesized, one has to wait for financial data to be produced in the real world before using them in model training and validation. The third challenge with financial data is the high level of noise associated with highfrequency trading data. Since high-frequency data in finance are invariably associated with a high level of noise, the machine learning models trained on such noisy data are intrinsically imprecise. Data evolution in finance poses the fourth challenge. Unlike data in most other fields, where the data features do not change with time, the features in financial data in harmony with financial markets evolve and change with time. This implies that financial variables will not carry the same meaning and significance over a long period, say one decade. The changes in the semantic meaning and significance of financial variables make it particularly difficult for the machine learning model to derive a consistent explanation and learning over a reasonably long period [27].

**Black-box nature of the models:** Machine learning and artificial intelligencebased models are black-box in nature [28]. In these models, while the outputs from the model are available and most of the time, they are easily interpretable, the biggest shortcoming is their lack of power of explanation of the output. In many critical applications in finance, mere outputs are not sufficient, and strong logical support for explaining the output is mandatory to instill sufficient confidence in the minds of the decision-makers. In absence of such explainable characteristics of the machine learning models, it will always remain a difficult job for the modelers to advocate the suitability of such models in critical business use cases.

**Validation challenges of the models:** Due to their higher complexity and opaqueness in operation, the machine learning models pose some significant challenges to risk management and validation [29]. While the regulators demand the machine learning models to comply with the SR 11–7 and OCC 2011–2012 standards of risk management, the optimum execution of the models may not be possible if all those guidelines are to be strictly adhered to. Model risk is an event that occurs when a model is designed following its intended objective but it introduces errors while in execution yielding inaccurate results from the perspective of its design and business use case. Another manifestation of model risk can happen when a model is built and deployed inaccurately or with faults without proper knowledge about its limitations and shortcoming.

**Challenges in model testing and outcome analysis:** The performance of a model and its accuracy in testing are evaluated by outcome analysis [29]. Since the neural network model has a natural tendency to overfit or underfit the data based on the training, it is imperative on the part of the model evaluation team to address the bias-variance trade-offs in the training and validation process. Since the traditional k cross-validation procedure used in backtesting of predictive models does not work effectively for machine learning model validation, the machine learning model validators should take particular care in carrying out normalization and feature selection before model training and validation. Validation loss and accuracy should be strictly

*Introductory Chapter: Machine Learning in Finance-Emerging Trends and Challenges DOI: http://dx.doi.org/10.5772/intechopen.101120*

monitored and analyzed to ensure that no overfitting or underfitting is present in the model before it is sent to the production phase. Neural network models are also difficult to evaluate on their sensitivity analysis as these models lack explainability. Since establishing a functional relationship between the explanatory variables and the target variable in a neural network model is a difficult task unlike statistical models, sensitivity analysis between the inputs and the output may involve a computationally involved challenge while the results of the analysis may be quite complex.

**Challenges with models designed by vendors:** As per the requirements specified in SR 11–7 and OCC 2011–2012 standards, models supplied by vendors are required to adhere to the same rigor as internally developed models [29]. However, in many practical situations, due to proprietary restrictions testing of vendors supplied models becomes a challenge. For vendor-supplied models, banks and financial institutions will have to mostly rely on softer forms of validation. The softer form of validation may include periodic review of model performance and conceptual soundness, stringent assessment of model customization, review of the development process, and applicability of the model in the portfolio of operations of the bank.
