Comprehensive Notes on Agentic Decision-Making and Market Modeling at Fetcher

Fetcherr

Definition and Function of Probabilistic Voxels: Unlike visual voxels used in traditional imaging, these are units representing the probability of specific market events. These combinations include attributes of the product class and the customer class. Events modeled within these voxels include:
- The occurrence of a transaction.
- The event of a cancellation.
- Competitor repricing actions.
The Agentic Workflow Sequence: The market models function as tools within a broader agentic workflow designed for actual decision-making. The process follows a specific cycle:
1. Data Consolidation: Gathering information from disparate sources to feed the model.
2. Training and Prediction: Once trained, the market model predicts future market dynamics.
3. Scenario Simulation: Predictions are utilized to simulate various future scenarios.
4. Decision Injection: A specific scenario is selected to inject the optimal decision into enterprise systems at every critical point.

Proprietary Enterprise Data: The foundation of the model is built on internal historical transactions unique to the organization.
External Predictive Data: To enhance the market model's accuracy, data is enriched with external information that can be predictive of market shifts, including:
- Real-world events.
- Capital markets data.
- Weather patterns.
- Any information type that correlates with market behavior.

Deep Learning Systems: The market model is a deep learning system that allows for an examination of its internal logic, similar to language systems. It utilizes attentional layers to determine which factors are most influential at any given time.
The New Competitor Case Study: In a specific application involving a new competitor entering the market mid-year, the model used its attentional layers to analyze the entrance.
- The new competitor was represented visually in yellow.
- The model demonstrated the ability to redistribute its "attention" among different competitors according to their calculated importance, based on the specific attributes of the new entry versus existing market players.

Human Visualization Constraints: While the market model's output is multidimensional, humans typically visualize between $2$ and $4$ dimensions. This requires engineers to "slice" the multi-dimensional cube at specific attributes, such as a specific product, customer type, transaction time, or delivery time.
Demand Surface Variables: A common visualization involves playing with specific dimensions on a coordinate system:
- X-axis: The "own price," or the price assigned to the enterprise's product.
- Y-axis: The price chosen by a competitor for a competing alternative product.
- Resulting Surface: The expected demand as quantified by the market model based on historical data.
Visual Interpretation of Demand Surfaces:
- Blue Areas: Represent zones where expected demand is significantly low, typically occurring where the enterprise price is substantially higher than the competitor's price.
- Red Areas: Represent high-demand zones.
- Equilibrium Line: The boundary where the public appears indifferent to the price difference between the two products.

Comparative Market Slicing: By slicing the data cube in different locations, the system identifies differences between distinct markets:
- Low Impact Markets: Scenarios where competition has a minimal effect on demand.
- Perceived Value Shifts: A case where the equilibrium line moves to the right by $50\,\text{price units}$ indicates that the perceived value of the product is higher than that of the competition in that specific market.
- High Contrast/Competitive Markets: Markets where the demand surface has higher contrast, indicating that every cent matters in pricing actions.
Temporal Correlations (Airline Pricing Example): The demand surface changes over time in response to external variables. For instance, in airline pricing, demand surfaces fluctuate in correlation with fuel prices. A documented case showed demand being impacted as fuel prices rose around April of a given year.

Digital Twin Simulation: Predictions are fed into a digital twin of the business arena to evaluate the expected reward for any specific pricing policy.
Reward Landscapes and Decision Hill-Climbing: The decision policy can be simplified as a two-parameter policy (Decision 1 and Decision 2). Every combination results in an expected reward (e.g., revenue).
- The system seeks to stay at the "peak of the hill" in the reward landscape.
- Because reality changes constantly, the landscape also shifts.
Decision-Making Regimes: The landscape contains multiple "hidden hills," each representing a different decision-making regime. If external reality changes such that the hill of "Regime B" becomes higher than the hill of "Regime A," the system automatically moves the decision policy to the more profitable regime.

AB Testing Methodology: Systems are deployed on a subset of markets (Target Group) and compared against a correlated Control Group.
Revenue Uplift Statistics: Successive testing of these quantitative decision-making systems demonstrates statistically significant results. On average, the revenue distribution is shifted to the right, showing a revenue uplift of approximately $7\%$ .
Transitioning the Industry: These tools facilitate the move from general "harness engineering" practices to truly reliable, enterprise-grade agent operations. Further technical details regarding the Large Market Model and Machine Learning are handled by specialists like Hadar Sharbit, the VP of Large Market Model and Machine Learning at Fetcher.", "title": "Comprehensive Notes on Agentic Decision-Making and Market Modeling at Fetcher"}