This paper critically evaluates the long-term robustness of LLM-based financial investing strategies by introducing FINSABER, a backtesting framework designed to mitigate survivorship, look-ahead, and data-snooping biases. The framework integrates multi-source data (prices, news, filings) from 2000 to 2024, includes delisted stocks, and employs a two-step pipeline combining selection-based and timing-based strategies over rolling windows. The authors replicate prior evaluations on select symbols (TSLA, NFLX, AMZN, MSFT) and short periods, finding that LLM advantages are inconsistent and often matched by traditional benchmarks. Extending the evaluation to two decades reveals that Buy and Hold and other simple strategies frequently outperform LLM agents. A broader composite evaluation over 91+ symbols shows that LLM strategies yield negative or near-zero Sharpe ratios, indicating poor risk-adjusted returns. Market regime analysis further demonstrates that LLM strategies are overly conservative in bull markets, missing upside, and overly aggressive in bear markets, incurring disproportionate losses. The findings underscore the need for regime-aware and adaptive risk management in LLM-based trading, rather than merely scaling model complexity. The paper provides a robust benchmarking framework and empirical evidence that current LLM investing strategies do not reliably outperform the market in the long run.