AI Investment Glossary
Essential terms and definitions for understanding AI investment bias, large language models, and machine learning in financial markets
CORE CONCEPTS
Hallucination Yield
The systematic biases that large language models (LLMs) exhibit when making investment recommendations. Named after AI "hallucinations" (confident but incorrect outputs), hallucination yield refers to the measurable premium or discount that AI models assign to certain stocks based on training data patterns rather than fundamental analysis.
Example: If multiple AI models consistently predict higher returns for Tesla compared to traditional automakers, this represents a positive hallucination yield for Tesla.
Related: Introduction to Hallucination Yield
AI Investment Bias
Systematic preferences or prejudices that artificial intelligence models display when analyzing investments. These biases emerge from training data patterns, algorithmic design choices, or inherent limitations in how AI processes financial information.
Types include: Frequency bias (overmentioning popular stocks), sentiment bias (consistently positive/negative tone), recency bias (overweighting recent events), and competitive bias (favoring/disfavoring rival companies).
Systematic Bias
Consistent, repeatable patterns of deviation from objective analysis across multiple AI models or time periods. Unlike random errors, systematic biases indicate underlying structural issues in AI training or methodology.
In AI investing: When ChatGPT, Claude, and Gemini all consistently overestimate crypto returns, this suggests systematic bias rather than coincidental error.
LARGE LANGUAGE MODELS
Large Language Model (LLM)
Advanced AI systems trained on vast amounts of text data to understand and generate human-like language. Examples include ChatGPT (GPT-4o), Claude Sonnet, Google Gemini, and others. In investment analysis, LLMs process financial information and generate recommendations based on learned patterns.
Key models we track: GPT-4o, Claude Sonnet-4, Gemini 2.0, DeepSeek-v3, Qwen-Max
Training Data
The massive collection of text from websites, books, articles, and other sources used to teach AI models language patterns. Training data quality and composition directly influence model biases and recommendations.
Investment impact: If training data contains more bullish Tesla articles than Ford articles, the AI may exhibit systematic bias toward Tesla in recommendations.
Training Data Cutoff
The point in time when an AI model's training data ends. Models cannot access information after their cutoff date, creating blind spots for recent market events, earnings, or news that could significantly impact investment analysis.
Example: A model with a 2023 cutoff wouldn't know about 2024 AI breakthroughs that affected NVIDIA's stock price.
AI Hallucination
When AI models generate confident-sounding but factually incorrect information. In investment contexts, this might include citing non-existent financial reports, inventing stock prices, or making up company partnerships.
Investment risk: An AI might confidently predict a stock merger that doesn't exist, leading to poor investment decisions.
Confidence Score
A numerical measure (typically 0-1 or 0-100) indicating how certain an AI model is about its predictions. Higher confidence doesn't necessarily mean higher accuracy, especially when models are systematically biased.
Our analysis: We track confidence levels to identify when models are uncertain, which often correlates with complex or controversial stocks like Tesla.
FINANCIAL ANALYSIS
Bullishness Score
A quantified measure of how optimistic an AI model is about a particular investment. We calculate bullishness based on language sentiment, predicted returns, and overall recommendation tone.
Scale: Typically 0-100, where 50 is neutral, above 50 is bullish, and below 50 is bearish.
Sentiment Analysis
The process of analyzing text to determine emotional tone, opinions, and attitudes. In AI investment research, we use sentiment analysis to quantify how positively or negatively models discuss different stocks.
Application: Measuring whether AI models use more positive language when discussing Apple versus Microsoft, indicating potential bias.
AI Consensus
The level of agreement among different AI models about an investment recommendation. High consensus suggests either strong fundamental factors or shared systematic biases.
Examples: Bitcoin shows unanimous AI bullishness, while Tesla creates maximum AI disagreement.
Return Prediction
AI-generated forecasts of expected investment performance over specific timeframes (3-month, 1-year, 5-year). These predictions help identify AI biases and compare model optimism levels.
Important: AI return predictions are for research purposes only and should never be used for actual investment decisions.
TYPES OF AI BIAS
Competitive Bias
When AI models developed by one company show systematic favoritism or prejudice against competitor companies. This occurs when corporate interests inadvertently influence model training or recommendations.
Example: Google's Gemini predicting significantly lower returns for Apple compared to other AI models.
Frequency Bias
The tendency for AI models to disproportionately mention or recommend stocks that appeared frequently in their training data, regardless of current investment merit.
Result: Popular media darlings like Tesla or NVIDIA may receive inflated attention compared to equally viable but less-discussed companies.
Recency Bias
Overweighting recent events or trends in investment analysis, even when they may not be representative of long-term patterns. AI models may extrapolate short-term movements into long-term predictions.
Impact: Models trained during crypto bull markets may exhibit persistent optimism about cryptocurrency investments.
Size Bias
The tendency for AI models to overestimate the market importance, size, or influence of certain companies based on their prominence in training data rather than actual financial metrics.
Example: An AI might overstate Tesla's market share in the automotive industry due to disproportionate media coverage.
RESEARCH METHODOLOGY
Standardized Prompts
Identical questions or requests sent to different AI models to ensure fair comparison. Using standardized prompts eliminates variables that could affect model responses and isolates systematic biases.
Our approach: All models receive identical investment analysis requests to enable direct comparison of their recommendations and biases.
Cross-Model Analysis
Comparative analysis across multiple AI models to identify consensus patterns, outliers, and systematic differences in investment recommendations.
Value: Single-model analysis can be misleading; cross-model comparison reveals broader AI behavior patterns and reduces individual model quirks.
Temporal Analysis
Tracking how AI investment recommendations change over time to identify evolving biases, model updates, or shifts in systematic patterns.
Application: Monitoring whether AI models become more or less bullish on specific sectors as market conditions change.