Generative AI vs Discriminative AI: Glamour Boy vs Workhorse

Generative AI models like ChatGPT have been the talk of the town, captivating the world with their ability to produce creative content. However, in the realm of financial market analysis, it is the less glamorous but highly efficient discriminative AI models that matter more. In this article, I explore the differences between generative and discriminative AI models and shed light on the critical importance of discriminative models for analysts and investors.

Generative vs Discriminative AI Models

Generative AI and discriminative AI are the major types of AI models and serve distinct purposes. Generative AI models are designed to generate new data based on the training data they were exposed to. They learn the patterns and characteristics within the data and can produce new content based on those patterns.

Over the last few months, we have all witnessed the meteoric rise of ChatGPT, a SaaS tool based on the generative AI model GPT (currently in its 4th version). Other popular generative AI tools, which happily piggyback on ChatGPT’s fame, are Google Bard, Microsoft Bing Chat, and Claude 2 by Anthropic.

Discriminative AI models are distinctly different from generative AI in terms of their structure and use cases. These models focus on classification tasks. Their primary objective is to categorise input data into specific classes or categories. Discriminative models learn to distinguish between different classes based on relevant features present in the data. Discriminative models have demonstrated high accuracy in tasks like sentiment analysis, fraud detection, and image recognition.

Leading discriminative AI models like BERT and RoBERTa excel at classifying existing data into specific categories without generating new content. These models focus on making precise distinctions, which is exactly what is required by software that needs a high degree of precision and accuracy. For instance, discriminative AI models are used by our social sentiment analytics tool, PUMP, to correctly classify social media posts mentioning financial assets into bearish, bullish, and neutral categories. Some AI-based market analysis and trading software packages also use discriminative AI models extensively.

Merits of Discriminative AI Models

Though generative AI models like ChatGPT have stolen the show, it is the discriminative models that matter the most for anyone involved in the field of investment and finance. These models work in the background to deliver the most precise estimates in social analytics tools and trading software.

Here are some of the key ways in which discriminative AI models outshine the overhyped generative AI team:

1. Higher Accuracy: Discriminative models offer superior accuracy compared to generative models. Precise classification is essential in financial markets, where data analysis drives investment decisions.

2. No Risk of "Hallucinations": Generative models are prone to “hallucinations”, producing fabricated or nonsensical content. In some cases, these fabrications might not be easy to spot, which makes the use of generative AI in anything that requires precision and facts highly problematic. Discriminative models, being purely focused on classification, do not suffer from this drawback.

3. High Predictive Power: Discriminative AI models capture patterns and relationships within data, resulting in high predictive power. Accurate predictions based on historical market data are invaluable for shaping effective investment strategies for anyone relying on AI to generate market insights.

4. Efficient Resource Utilisation: Discriminative models achieve comparable performance with modest resources, making them suitable for real-time applications in fast-paced environments.

5. Foundation for AI-based Financial Tools: Discriminative models serve as the backbone for various AI-based tools used in financial markets. The industry has so far not figured out a coherent and useful way of using generative AI models in financial software. This is perhaps an area for further active R&D by AI researchers and developers.

6. Explainability: Discriminative models are generally more interpretable and explainable, allowing analysts and investors to understand the rationale behind AI-driven decisions, something that is critical for fostering trust in the technology. The mechanism behind the classificatory algorithms of these models is clear-cut and transparent. In contrast, generative AI models are akin to “black boxes”, often generating perplexing results that are hard to explain.

With higher accuracy, no hallucination risks, high predictive power, and efficient resource utilisation, discriminative models are the true workhorses within AI software packages used in the financial world. Let the fiction writers pedestalise generative AI; discriminative models are where the biggest value of AI technology lies for financial analysts, investors, and traders.