This is the third in a series of four articles detailing Transparently.ai’s approach to harnessing artificial intelligence and machine learning to build its tools to detect accounting fraud.
We have discussed the general approach employed by Transparently.ai to assess potential accounting manipulation risk. We have introduced the 14 risk clusters that form the basis for identifying patterns of risk. We now explain how these patterns are combined to generate a quantitative risk score for each company, known as the Transparently Risk Score.
Deriving a score for each risk cluster
We've discussed that each cluster looks for specific patterns. What our AI does first is take all those different patterns and clues within a single cluster (like “Investing Activity” or “Income Quality”) and boil them down into one clear score for that cluster.
How does it do that? Within each of the 14 risk clusters, there are many individual pieces of evidence or clues – these are what we call factors. For example, take "Investing Activity.” One factor in this cluster might be "unusually high spending on intangible assets," another might be "a lot of undisclosed transactions," and a third could be "a sudden sale of investment products."
These clues or factors are the elements that our AI considers to generate one clear score for that cluster. This is the process that it uses:
- Evaluating each clue (Factor): For every single clue (factor) within a cluster, our AI first assesses its individual significance. It's like a detective looking at a piece of evidence and deciding, "How strong is this clue? Is it a minor detail or a major red flag?" This assessment is based on how often that specific clue has been present in past cases of manipulation versus legitimate financial activity. So, a very strong, consistent red flag gets a higher individual "score" or weight.
- Comparison to the norm: The AI doesn't just look at the clue in isolation. It compares it to what's normal for companies of similar size, in the same industry, and at that point in time. If "TechGadget Co." has high intangible asset spending, but all other tech companies do too, it's less of a red flag. But if "TechGadget Co." is an outlier, that clue becomes much more significant.
- Combining the evidence: Once each clue has been evaluated and weighted based on its significance and deviation from the norm, the AI then combines all these individual assessments for that specific cluster. It's not just a simple average; it's a sophisticated calculation that considers the interplay between the clues. Some clues might amplify each other, while others might be less critical when seen alongside stronger evidence.
- Generating the cluster score: The result of this combination is a single, numerical score for that entire risk cluster. This score represents the overall level of concern or risk identified within that specific area of the company's financial behavior. A higher score means more and stronger red flags were found within that cluster, indicating a greater likelihood of potential manipulation in that particular aspect of the company's operations.
So, for each cluster, the AI acts like a judge, weighing all the evidence presented by the individual factors, considering their context, and then delivering a verdict in the form of a single, comprehensive score.
Combining the cluster scores to create the overall 'Transparently Risk Score'
Once we have a score for each of the 14 individual risk clusters, the Transparently tool then combines them into a single, overall 'Transparently Risk Score' through a sophisticated process that is akin to a master chef blending ingredients into a complex dish.
Here's how our AI weighs and combines these different aspects of risk:
- Dynamic weighting based on context: Not all risk clusters are equally important for every company, or at every point in time. For instance, "Investing Activity" might be a huge red flag for a rapidly expanding tech company, but less critical for a mature utility. Our AI doesn't apply a one-size-fits-all weighting. Instead, it dynamically adjusts the importance (or 'weight') of each cluster based on:
- Industry: What's normal and risky for a bank is different from a retail chain.
- Company size and stage: A startup's risk profile differs from a multinational corporation.
- Economic environment: Certain risks become more pronounced during economic downturns.
- Interplay of clusters: If "Business Manipulation" is high, it might make "Income Quality" even more critical, as one problem could be exacerbating the other.
- Non-linear relationships: The relationship between different risks isn't always straightforward. A high score in one cluster might be manageable on its own, but when combined with a high score in another, it could create an exponentially higher overall risk. Our AI models are designed to understand these complex, non-linear interactions, recognizing that the sum of the parts can be much greater than the individual components. It's like how certain chemicals are harmless alone but explosive when mixed.
- Historical performance and predictive power: The AI continuously learns from past data. It observes which combinations and magnitudes of cluster scores have historically led to actual accounting scandals or financial distress. This historical learning helps it refine the weighting and combination logic, giving more emphasis to clusters or combinations that have proven to be strong predictors of future problems.
- Proprietary algorithms: At its core, the mathematical combination of risk factors is driven by proprietary algorithms that have been developed and refined over years. These algorithms are designed to mimic the nuanced judgment of an expert forensic accountant, but with the ability to process vastly more data and identify patterns beyond human capacity.
So, the overall 'Transparently Risk Score' isn't just an arithmetic sum. It's a carefully calculated, dynamically weighted, and historically informed synthesis of all 14 cluster scores, designed to provide the most accurate and predictive measure of a company's accounting manipulation risk.
The difference between the Risk Score and the Risk Rating
Finally, what's the difference between the 'Transparently Risk Score' (which is a number, like 45.69%) and the 'Risk Rating' (which is a letter grade, like A+ or F)? Why do we have both, and what does each tell a user about a company's accounting risk?
The Transparently Risk Score
This is the precise, granular measurement of a company's accounting manipulation risk. Imagine it like a thermometer reading. It gives you the exact temperature of the risk, down to decimal points. It's a continuous scale from 0% (very low risk) to 100% (extreme risk).
It tells a user the exact magnitude and intensity of the risk. It's incredibly useful for tracking subtle changes over time, comparing companies with very similar risk profiles, and understanding the precise degree of concern. If a company's Score moves from 45.69% to 46.10%, that small shift might indicate a slight worsening trend that a broader category wouldn't capture. It's the quantitative detail.
The Risk Rating
This is a categorical, simplified interpretation of the numerical Score. Think of it like a doctor telling you if you have a 'normal temperature,' a 'mild fever,' or a 'high fever.' It groups ranges of the numerical Score into easily understandable letter grades, from A+ (lowest risk) to F (highest risk).
It provides a quick, intuitive understanding of the severity or category of the risk. It's designed for rapid assessment and easy communication. If you see an 'F' rating, you immediately know the company is in the highest risk category, without needing to interpret a specific percentage. It's a qualitative summary.
Why we have both
We use both the Score and the Rating because they serve complementary purposes:
- Precision versus simplicity: The Score offers the precision needed for detailed analysis, trend tracking, and sophisticated modeling by financial professionals. The Rating offers simplicity and immediate comprehension for quick decision-making, portfolio screening, and broader communication.
- Context and actionability: The rating provides immediate context for the Score. A 70% risk score might sound high, but an 'F' rating instantly confirms that it falls into the most severe category, prompting immediate attention. Conversely, if two companies both have a 'C' rating, their specific Score scores (e.g., 51% vs. 59%) can show which one is closer to moving into a higher risk category.
In essence, the Transparently Risk Score gives you the exact data point for deep dives and trend analysis, while the Risk Rating provides a clear, actionable benchmark for quick understanding and comparative assessment. These are the key outputs of the Transparently Risk Engine that underpin the value of our tools to the asset managers, auditors and regulators that use them.
