QuantCode-Bench is a benchmark designed to evaluate large language models on generating executable algorithmic trading strategies from natural language descriptions using the Backtrader framework. The benchmark comprises 400 tasks sourced from Reddit, TradingView, StackExchange, GitHub, and synthetic data, categorized by difficulty. Evaluation follows a four-stage pipeline: syntactic correctness (Compilation), successful backtest execution (Backtest), presence of at least one trade (Trade), and semantic alignment with the task description via an LLM judge (Judge). This layered approach reveals that while frontier models achieve near-perfect compilation, they struggle with correct operationalization of trading logic, proper API usage, and adherence to task semantics, with the best single-turn Judge Pass around 75.8% (claude-opus-4.6). In an agentic multi-turn setting with iterative feedback, performance improves dramatically, reaching 95-98% Judge Pass for top models, indicating that many errors are locally repairable. Error analysis shows that the most common failures are strategies that compile and backtest but generate no trades (17.8%) and incorrect handling of Backtrader line objects (13.1%). The benchmark highlights that trading strategy generation is a distinct domain-specific code generation task requiring not only technical correctness but also deep financial logic and semantic understanding. QuantCode-Bench is publicly released to foster research in domain-specific code generation and agentic software repair in finance.