Israeli Startup 42AI Raises $62.5 Million to Deploy Behavioral World Model in Finance
42AI has emerged from stealth with a $62.5 million seed round and a transformer‑based model that predicts trader behavior.
Stealth Exit and Funding
AI firms are increasingly blending behavioral science with large‑scale models to gain an edge in financial markets.
Israeli startup 42AI announced it has closed a $62.5 million seed round and is now operating publicly after three years of development. The financing will support the rollout of its “behavioral world model,” which the company says can predict how individual traders will react to volatile market conditions.
How the Model Works
CEO and co‑founder Amit Gal‑Or described behavior as “probabilistic” rather than deterministic, arguing that underlying psychological tendencies persist across situations. Co‑founder and CTO Ehud Ezra added that a person’s tendencies “will always play out in certain areas to a probability or distribution.”
Instead of relying solely on past actions, 42AI claims its model examines how people respond to a variety of seemingly unrelated events—such as completing KYC processes, onboarding steps, support calls, money transfers, and actual trades. Ezra explained that the system captures sentiment, voice cues, call frequency, and other interaction metrics, then feeds these data points into a transformer‑based world model.
The model converts observations into mathematical representations of traits, which are used to generate probability distributions for possible decisions under different circumstances. A separate language model produces explanations for each prediction.
Training Data and Validation
According to Gal‑Or, the team first gathered psychometric assessments and behavioral scenario responses from thousands of participants, then incorporated government and biobank datasets covering more than two million people. Customer datasets now span “almost 30 million people.”
In live deployments at a handful of brokerages, the system analyzes existing records—transactions, support interactions, and other activity—without requiring personally identifiable information. Gal‑Or emphasized that “the model never knows who the person is,” even though it creates individual behavioral profiles from linked activity.
Performance Metrics
When applied to market‑trade prediction, 42AI reports that its model reaches 98% accuracy in terms of whether trading flows fall within a predicted range. The company also cites an F1 score between 0.8 and 0.85 for forecasting trade direction and volume, indicating a balance between correct predictions and false alarms.
Ezra noted that the model is trained to preserve a person’s underlying tendencies; if a typically risk‑averse individual is simulated as a risk‑taker, the training reward penalizes the outcome. He also warned that accuracy drops when the system lacks critical situational context, such as advice from a friend influencing a trader’s choice.
Limitations and Outlook
Gal‑Or acknowledged a “natural ceiling” because the model cannot access every experience that shapes a person’s decisions. The current implementation remains confined to the financial‑trading environment, where the necessary data streams are available.
Despite these constraints, the company believes its approach—merging psychology, neuroscience, genetics, and AI—offers a new way to anticipate human behavior in high‑stakes settings.
Whether this blend of behavioral science and transformer models will reshape trading strategies remains to be seen.
