Gold/Silver Ratio: GenAI with Quant Agent on Azure AI Foundry

The article”Gold/Silver Ratio: GenAI with Quant Agent on Azure AI Foundry” was originally posted on DataGeeek.

1. Introduction: The Strategic Edge of Agentic Finance

In the contemporary landscape of quantitative finance, the bottleneck is no longer data availability, but the speed of insight generation. Leveraging the Microsoft AI Foundry ecosystem, we have moved beyond static scripting into the realm of Autonomous Financial Agents. This article explores how a specialized agent can navigate precious metal volatility by analyzing the Gold/Silver ratio with high-performance precision.

2. Infrastructure: Model Deployment on Microsoft AI Foundry

The intelligence behind this analysis is not a local script but a deployed model instance on Microsoft AI Foundry. We utilize the GPT-4o model, deployed as a scalable web service within the Foundry environment.

• Endpoint Security: By using the Azure OpenAI service within AI Foundry, we ensure that financial queries and data remain within a secure, enterprise-grade perimeter.
• Agentic Logic: The “Agent” is more than just a model; it is a programmed entity with a specific System Prompt that defines its persona as a Quantitative Researcher. This allows the model to “reason” through the necessary steps: from library loading to data merging and final visualization.

3. The Technical Bridge: Python-R Integration

One of the most powerful features of our AI Foundry Agent is its multi-lingual capability. It bridges the gap between Python and R using the rpy2 library, creating a high-performance research pipeline.

The R Ecosystem in Play:

• tidyquant & timetk: These packages are the engine for our time-series analysis. tidyquant handles the seamless fetching of GC=F and SI=F data, while timetk manages the complex task of plotting with built-in smoothing algorithms.
• dplyr & lubridate: Essential for the “tidy” manipulation of data, allowing the Agent to perform inner_join operations and date-based filtering with surgical precision.

4. Methodology: Taming the Noise with Visual Precision

To extract actionable trends, the Agent is instructed to apply a LOESS smoothing algorithm. By strictly setting .line_size = 1.5 and .smooth_size = 1.5, we ensure the trendline is bold enough to be the primary focus for analysts, effectively “taming” the daily price volatility.

5. Conclusion: Scaling Quantitative Research

The synergy between Microsoft AI Foundry, deployed LLMs, and specialized R packages represents the future of financial research. We have replaced manual data wrangling with an autonomous, standardized agent that can be scaled across thousands of different asset pairs with a single command.

The ABI Connection (Bridging Python to R in VS Code)

For the script to run locally in VS Code, we must establish a robust Application Binary Interface (ABI) connection. This is handled by the rpy2 library, which serves as the translation layer between Python and the R interpreter.

• Synchronization: The script uses a localconverter to transform Python data types into R objects in real-time.
• Environment Sync: Before the Agent’s code is executed, the script automatically synchronizes the working directory (setwd) so that files generated by R (like the ratio_plot.png) are immediately accessible to the Python environment for rendering.

import os
 
# Force rpy2 to use ABI mode to avoid the Windows CFFI conflict
os.environ['RPY2_CFFI_MODE'] = 'ABI'
 
import rpy2.robjects as robjects
from rpy2.robjects.packages import importr
print("Interface initialized in ABI mode.")

The Integrated Agent Script:

import os
import httpx
from openai import AzureOpenAI
import rpy2.robjects as robjects
from rpy2.robjects import pandas2ri
from rpy2.robjects.conversion import localconverter
from IPython.display import Image, display
 
#Microsoft AI Foundry - Azure OpenAI Connection
client = AzureOpenAI(
    api_version="2024-12-01-preview",
    azure_endpoint="AZURE_OPENAI_ENDPOINT",
    api_key="AZURE_OPENAI_KEY", 
    http_client=httpx.Client(verify=False, trust_env=False)
)
 
def run_updated_agent(user_request):
    system_instructions = (
        "You are a Quantitative Researcher. MANDATORY: All output, comments, and labels in English. "
        "Strict Operational Guidelines:\n"
        "1. Libraries: library(tidyquant), library(timetk), library(lubridate), library(dplyr), library(ggplot2).\n"
        "2. Analysis: Fetch GC=F and SI=F for 3 years, merge via inner_join, and calculate 'ratio'.\n"
        "3. Visualization: Use timetk::plot_time_series with .interactive = FALSE and .title = \"Gold/Silver Ratio\".\n"
        "4. Precision: Set .line_size = 2 and ALWAYS set .smooth_size = 2 for the smoothing line.\n"
        "5. Set title font face and axis texts font face to 'bold', and size to 16 with theme() function.\n"
        "6. EXPORT: Save using 'ggsave(\"ratio_plot.png\", width = 10, height = 6)'.\n"
        "7. Output ONLY raw R code."
    )
 
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[
            {"role": "system", "content": system_instructions},
            {"role": "user", "content": user_request}
        ]
    )
     
    # Cleaning any markdown or headers to get raw code
    agent_code = response.choices[0].message.content.strip()
    if agent_code.startswith("```"):
        agent_code = "\n".join(agent_code.split("\n")[1:-1])
     
    print("-" * 40)
    print(agent_code)
    print("-" * 40)
     
    try:
        with localconverter(robjects.default_converter + pandas2ri.converter):
            # Synchronize working directory
            robjects.r(f'setwd("{os.getcwd().replace("\\", "/")}")')
            robjects.r(agent_code)
             
            if os.path.exists("ratio_plot.png"):
                display(Image(filename="ratio_plot.png"))
    except Exception as e:
        print(f"Agent Error: {e}")
 
# Execution
run_updated_agent("Plot the Gold/Silver ratio for the last 3 years with a smooth line.")

Source: Yahoo Finance

NOTE: This article was generated with the support of an AI assistant. The final content and structure were reviewed and approved by the author.