the-house

The House Problem

Overview

This project explores a competitive market-making environment inspired by real-world electronic trading systems.

Participants act as market makers in a simulated betting exchange where each round contains a hidden payout value. Every participant receives a noisy estimate of the true value and must continuously quote bid and ask prices while competing against multiple fixed-strategy operators and informed traders.

The objective is to maximize long-run expected profit while balancing:

• spread competitiveness
• fill probability
• adverse selection risk
• informed trader exploitation

The project combines ideas from:

• market microstructure
• quantitative trading
• probabilistic modeling
• simulation systems
• reinforcement learning environments
• adversarial multi-agent systems

---

Problem Statement

For each round:

• A hidden value V ∈ [0,1000] is generated.
• Every market maker receives a noisy signal:

$$s = V + \epsilon$$

where:

$$\epsilon \sim \mathcal{N}(0,50^2)$$

Each participant must quote:

• Bid → price willing to buy at
• Ask → price willing to sell at

Punters and informed traders then choose the best available prices across all active booths.

The challenge is to design a pricing strategy that captures uninformed flow while minimizing losses against perfectly informed traders.

---

Strategy Design

The final strategy was developed through:

• historical data analysis
• operator reverse-engineering
• fill competition modeling
• spread optimization
• simulation-based evaluation
• parameter search

The system dynamically adjusts spreads based on signal extremity and competitive pressure rather than using fixed-width pricing.

Core ideas implemented:

• tighter spreads in high-liquidity regions
• wider spreads near edge signals
• asymmetric protection against toxic buy-side flow
• controlled aggressiveness to avoid adverse selection

---

Repository Structure

the-house-problem/
│
├── auction_history.csv        # Historical market data
├── final_submission.csv       # Final optimized pricing table
├── house.pdf                  # Competition/problem statement
├── Untitled.ipynb             # Research + simulation notebook
├── plots/                     # Generated visualizations
└── README.md

---

Data Analysis

The historical dataset contains:

• true payout values
• noisy market maker signals
• quotes from multiple house operators
• fills received
• realized PnL

Each operator exhibited distinct behavior patterns:

Operator	Behaviour
Tight	Narrow spreads, high fill rate
Wide	Conservative pricing
Skewed	Directionally biased quotes
Noisy	High quote variance
Fade	Counter-signal positioning

By reconstructing their quoting behavior, it became possible to identify pricing gaps and profitable spread regions.

---

Simulation Framework

A custom simulator was built to evaluate strategies under realistic competition dynamics.

The simulator models:

• competing market makers
• best-price fill logic
• informed trader behavior
• adverse selection
• spread competition
• round-level PnL

The framework enables rapid backtesting across thousands of simulated rounds.

---

Optimization Approach

Instead of manually choosing quotes, the strategy was parameterized and optimized systematically.

Optimization variables included:

• base spread width
• edge widening factor
• directional skew
• aggressiveness level

A large-scale parameter sweep was then used to identify profitable configurations under varying market conditions.

---

Key Concepts Explored

• Market Making
• Bid-Ask Spread Dynamics
• Adverse Selection
• Liquidity Provision
• Quantitative Pricing
• Simulation-Based Optimization
• Competitive Multi-Agent Systems

---

Example Quote Structure

The strategy generates quotes of the form:

$$bid = s - \frac{spread(s)}{2}$$ $$ask = s + \frac{spread(s)}{2} + skew$$

where spread dynamically changes depending on market conditions and signal confidence.

---

Results

The final submission strategy achieved:

• competitive fill capture
• reduced toxic flow exposure
• stable long-run expected PnL
• improved robustness against informed traders

The project demonstrates how relatively simple probabilistic systems can produce complex emergent market behavior.

---

Technologies Used

• Python
• NumPy
• Pandas
• Matplotlib
• Jupyter Notebook

---

Running the Project

Clone the repository:

git clone https://github.com/your-username/the-house-problem.git

Install dependencies:

pip install pandas numpy matplotlib

Launch notebook:

jupyter notebook

---

Future Improvements

Potential extensions include:

• reinforcement learning market makers
• probabilistic fill modeling
• Bayesian value estimation
• multi-agent self-play
• dynamic spread adaptation
• learned adverse-selection detection

---

Author

Harshita Pulavarti

Built as an exploration of competitive market-making systems, simulation-driven strategy optimization, and adversarial pricing dynamics.