Avoid machine learning mistakes to unlock powerful new insights

Avoid machine learning mistakes to unlock powerful new insights

Written by Yuxin Yang, Practice Manager, Machine Learning, TensorIoT

While the business world has quickly embraced artificial intelligence (AI) and machine learning (ML), a subfield of AI, the financial services industry struggles with maximizing the advantages offered by the new technology. In its 2023 Financial Services GenAI Survey, Ernst & Young reports that 99 percent of respondents’ organizations are deploying AI, yet 20 percent express concern that their companies are not well-positioned to take advantage of the many benefits. For example, ML excels at analyzing enormous amounts of data, quickly identifying hidden patterns, and providing insights that become more accurate and detailed over time. These benefits can lead to better risk mitigation, more precise identification of customer needs and preferences, faster document processing and loan underwriting, and improved customer service.

Instead of seeing the benefits of improved data analysis, many financial organizations are experiencing counterproductive effects. They are wasting money on technology that doesn’t deliver as promised. They are getting low-quality data and inaccurate analysis that lead to poor decision-making. They are seeing workflows expand and grow more frustrating when workflows do not become more streamlined and frustration-free. They are wasting valuable hours on training for a system that quickly loses employee buy-in. For these reasons, it is essential for financial companies to understand the four most common mistakes businesses make when using ML and how to avoid them.

Four common pitfalls to avoid when using ML

While there are a variety of mistakes that can be made when a financial services organization implements ML into its operations, there are four that are made most often. These common pitfalls can erase the benefits ML can produce and prevent companies from keeping pace with or surging past the competition in their marketplace. These common mistakes include:

1. Working with low-quality data. ML is only as good as the data it is analyzing. Good data leads to good insights, which then is used to make good business decisions. Inaccurate or incomplete data will lead to poor decisions that negatively impact a business’s operations and financial stability and may even damage customer trust. 
2. Not properly managing ML system performance. Not monitoring ML systems can lead to processing errors and data inaccuracies that can negatively impact performance. If these issues aren’t corrected quickly, they can lead to long-term problems, including damage to a business’s reputation, regulatory non-compliance, poor decision-making and staff distrust, discriminatory practices, and financial losses.
3. Not creating adequate documentation. Detailed and organized documentation is essential for a successful ML system. Adequate record-keeping helps guarantee regulatory compliance, knowledge sharing, and internal oversight. Documentation also ensures more comprehensive, effective training, easier system reproduction, and simplifies future maintenance and system troubleshooting efforts.
4. Failing to foster a collaborative work environment. By not encouraging knowledge-sharing and a collaborative work environment, businesses are apt to develop departmental silos that foster inefficiency—such as when teams, unaware of what others are doing, perform redundant work. Poor communication also leads to a lack of new ideas, from varied staff working more closely together.

Avoid these pitfalls with low-code tools

Traditional ML solutions require the involvement of data scientists and software developers, increasing investments of time, costs, and other resources for financial institutions. Fortunately, low- or no-code platforms break down these barriers and enable businesses to create ML models via user-friendly interfaces.

Low- and no-code tools—such as Appian, Creatio, ZoHo Creator, Retool, Caspio, and Amazon SageMaker Canvas—are designed to reduce hand-coding so that ML models can be delivered faster and easier. These tools boast graphical user interfaces that allow users to drag and drop features instead of creating complex code. Ananya Bhattacharyya, a strategic global product and business leader at Mastercard, says low-code tools not only help financial enterprises innovate faster, they also enable them to navigate data in siloed systems and incompatible formats. “Combining sophisticated data processing along with Automl in a resilient low- or no-code development environment enables enterprises to build end-to-end digital native solutions faster and with minimum technical debt,” Bhattacharyya says. With low- and no-code tools, less technical staff can significantly impact ML system development and performance.

Here are some of the ways that low- and no-code tools can resolve the four common pitfalls that are preventing many organizations from best utilizing ML:

1. Data quality. Low- and no-code tools allow users to implement a clear and optimized data collection and storage policy, conduct regular data cleansing and validation, and continuously monitor data quality—all without any coding experience. An example of the importance of maintaining good data quality is American Express. The company uses its AI/ML-powered fraud detection system to continuously sift through massive amounts of high-quality transaction data and flag suspicious activity. The company says that due to the system, it can now generate a fraud decision in milliseconds each time an American Express card is used. That fraud system protects over $1.2 trillion in transaction value each year.
2. ML performance. Low- and no-code tools allow users to make changes easily, streamlining the workflow process. These tools can also provide flexibility in using a variety of different foundation models (FMs), such as Claude 2, Amazon Titan, Jurassic-2, and others for generative AI applications. In addition, their flexibility extends to ML services that process tabular data (such as spreadsheets and databases), text, and computer vision images, which allow users to more quickly assess which ready-to-use models are most appropriate or cost-effective. Users can automate testing and reporting and take action to improve communication and knowledge-sharing without having to go through a long and tedious coding process. For example, a private equity firm recently used a low-code tool to integrate ML into its investment and research operations. The company reports that this integration improved decision-making, streamlined processes, and gained an advantage over its competition.
3. ML documentation. With low- and no-code tools, it is also easier to assemble documentation. Users can utilize built-in templates to capture key project information, track documentation updates, and generate detailed reports. A leading national insurance carrier in California processed thousands of insurance claims through manual data extraction until it switched to ML. The company reports that low-code tool usage has transformed a painstakingly slow, manual documentation process into an automated one that allows staff to quickly process documents and provide valuable information to customers much faster.
4. Collaboration. Low- and no-code tools help pave the way for improved collaboration and knowledge-sharing by enabling the creation of shared workspaces and data pipelines. They may also offer such features as role-based permissions, change tracking, and document version control. Piraeus Bank, a Greek multinational financial services company, reports that its use of low-code tools has allowed it to break down company silos and create a work environment “where innovation, the exchange of ideas, and creativity are supported” and collaboration is fostered. Deloitte, a leading financial consulting and advisory company, says that its use of low-code tools has unlocked greater collaboration and improved efficiency that have enhanced “the speed of development and deployment productivity by 30–40% across client-facing and internal projects.”

From textual coding to visual coding

Technology company Radixweb reports that low-code tools are being used by 77 percent of organizations, and they will be responsible for more than 65 percent of application development activity in 2024. Quandary Consulting Group supports those figures by saying four out of five businesses in the United States now use low-code tools. These tools are powerful ML aids. Their usage can speed up app development, deployment, and management while reducing the risk of coding errors and allowing less tech-savvy employees to be more involved.

Effective ML implementation also requires a collaborative and transparent environment that fosters trust and mitigates concerns about bias. Companies enjoy wider ML system adoption with low-code tools because the drag-and-drop ease of low-code tools enables non-tech users to contribute more easily. They may also eliminate some job-displacement concerns that typically accompany ML implementation since more employees can participate and add value to the new system. Low- and no-code tools represent a game-changing opportunity for the financial services industry, helping companies address and overcome common roadblocks to ML system success. “Low- and no-code tools enable organizations to experiment, test, and deploy scalable digital native AI applications by integrating advanced data processing within a robust and easy-to-use development environment,” Bhattacharyya says. “This removes the barrier of keeping up with the ever-evolving technology landscape.” By combining the strengths of AI and low- and no-code ML, financial services organizations can experience unparalleled efficiency, accessibility, and innovation. This synergy will usher in a new era of empowerment and competition in the financial sector, creating limitless opportunities for those who join the ML revolution.

About the Author:

Yuxin Yang is the practice manager of machine learning at TensorIoT where she builds cutting-edge solutions for clients with an emphasis on leveraging data science and machine learning. She holds a master’s degree in computer engineering from Stanford University and a bachelor’s degree in electrical and electronics engineering from Columbia University. TensorIoT is an AWS Advanced Tier Services Partner that enables digital transformation and greater sustainability for customers through IoT, AI/ML, data and analytics, and app modernization. For more information, visit tensoriot.com.