M.Sc. Financial Engineering Thesis
Institution: Karlsruhe Institute of Technology (KIT), HECTOR School of Engineering & Management, Germany
Author: Ratchatapong Lukmuang
To protect proprietary research and prevent alpha decay, live execution scripts, signal parameters, raw tick data, and fitted model weights are intentionally omitted from this public repository.
This repository serves as a high-level technical overview of the quantitative framework, methodology, and system architecture developed for the thesis.
👉 Read the full thesis write-up, econometric analyses, and empirical results on LinkedIn:
🔗 Inside My M.Sc. Thesis: Building Algo-Trading at 9.85 Sharpe via Statistical Arbitrage for $100–$1,000 Retail Accounts
All backtest metrics are reported net of real-world frictions, including a $0.04 round-trip commission per share/lot, variable bid-ask spreads, overnight swap avoidance, and market-close protections.
| Metric | Champion (IS) | SPY B&H (IS) | Champion (OOS) | SPY B&H (OOS) |
|---|---|---|---|---|
| Testing Window | Nov 1, 2023 – Oct 21, 2025 | Nov 1, 2023 – Oct 21, 2025 | Oct 21, 2025 – Mar 31, 2026 | Oct 21, 2025 – Mar 31, 2026 |
| Net PnL (%) | +16.15% | +58.88% | +6.06% | -5.02% |
| Sharpe Ratio | 4.87 | 1.61 | 9.85 | -0.84 |
| Sortino Ratio | 37.98 | 2.14 | 65.45 | -1.32 |
| Calmar Ratio | 293.27 | 2.13 | 126.03 | -0.53 |
| Max Drawdown | -0.06% | -27.67% | -0.05% | -9.40% |
- OOS Win Rate: 90.12% (165 trades)
- Risk of Ruin: 0% (Validated via 10,000-path Monte Carlo simulation)
- Account Size Target: $100 – $1,000 retail base capital
- Traded Pair:
SPY.US CFDvs.US500 CFD(Pepperstone MetaTrader 5 API). - Capital Base: Designed for $100 to $1,000 retail accounts (unaffected by US PDT rules via European CFD access).
- Friction Model: $0.04 round-trip commission per lot on
SPY.US(breakeven threshold ~0.40 points).
- Dataset: ~2.5 years of tick-level data stored in Parquet format.
-
Spike Filter: Purges tick anomalies exceeding
$5 \times$ the 99th percentile rolling bid-ask spread. - Resampling Grid: 5-second sampling grid chosen to balance temporal resolution against "Market Silence" (achieving 92–94% fidelity without introducing phantom signals).
- Dataset Split: Chronological 80/20 split (In-Sample: Nov 2023 – Oct 2025, Out-of-Sample: Oct 2025 – Mar 31, 2026).
Evaluated 7 econometric candidates across linear (EG-ECM), dynamic-linear (Kalman Filter + OU), and non-linear (Gaussian Copula) families. Model selection was driven by a novel evaluation metric:
- Champion Model:
Copula_OU (AR1)at 1-minute resolution (CTES: 761.43, mean half-life: ~30 seconds). - Rejected Models: GARCH (rejected by ARCH-LM test) and SETAR (rejected by Chow test) due to lack of empirical support for added complexity.
To capture mean-reversion windows of ~30 seconds without bottlenecking statistical calculations, execution is split into two asynchronous modules:
- The Brain (1-minute): Calculates heavy statistics (cointegration parameters, copula marginals, OU constants).
- The Action (5-second): Monitors price spreads at microstructure resolution (92–94% fidelity) and executes trades authorized by The Brain.
🔍 Click image to enlarge (Ctrl/Cmd + Click to open in new tab)
- Session Guard: Blocks entry signals during the first 90 minutes of the New York trading session to prevent losses caused by zero tail dependence violations, momentum jumps, and bid-ask spread noise at the open.
- Position Sizing: Continuous Kelly position sizing capped at a 70% maximum margin ceiling.
- Circuit Breakers: 10% equity drawdown halts new entries; 20% equity drawdown triggers hard liquidation.
- Swap & Close Protection: Halts entries 10 minutes prior to daily swap cutoff, force-closes positions 5 minutes prior, and resumes 3 minutes after.
Ratchatapong Lukmuang
M.Sc. Financial Engineering | Karlsruhe Institute of Technology (KIT)
Open to Quantitative Research, Quantitative Trading, and Quantitative Risk roles in Germany and internationally.
- LinkedIn Article: Inside My M.Sc. Thesis
- Contact Details: Available via LinkedIn profile