1,500 scalping scripts tested
One experiment gave Claude AI access to roughly 1,500 TradingView scalping strategies to see what a model would do with a huge library — and the video exposes a core point: more scripts ≠ more alpha. The creator’s premise — bulk-inspection to find unique logic, flag repainting indicators, and surface likely overfits — highlights that AI is useful for taxonomy and screening even if it doesn’t magically pick winners (youtube.com).
A trading YouTuber posted a simple stunt on April 6: give Claude access to roughly 1,500 TradingView scalping strategies and see whether a language model could find an edge that humans had missed. The video’s own description lays out the pitch in blunt terms. Claude was asked to analyze a large strategy library, pick the best systems, turn bots on or off, and manage a live portfolio through connected tools and APIs (youtube.com). That setup sounds like automated genius. It is really a stress test of a more ordinary idea: can an AI sort through a mountain of messy trading code faster than a person can. That distinction matters because TradingView strategies are not clean scientific objects. They are Pine Script files written by thousands of different people, with different assumptions, different timeframes, and different failure modes. TradingView’s own documentation says the platform’s Strategy Tester is built to simulate orders and show performance metrics for scripts declared as strategies, which makes it easy to generate a lot of backtest output quickly (tradingview.com). Easy testing is also why traders accumulate giant libraries of scripts in the first place. Once you can run thousands of variants, quantity starts to masquerade as evidence. The trap is older than AI. A strategy can look brilliant in historical data and still collapse the moment it meets live markets. TradingView warns that “repainting” is widespread, with more than 95 percent of indicators showing some form of historical-versus-realtime mismatch, and it explicitly calls future leaks and certain intrabar behaviors “unacceptable” because they can produce heavily misleading results (tradingview.com). In other words, a giant pile of scripts is not a giant pile of independent discoveries. It is often a giant pile of duplicated ideas, fragile parameter tweaks, and code that flatters itself in hindsight. That is where the Claude experiment becomes more interesting than its thumbnail. Anthropic’s current developer docs describe Claude as a model that can work through files, tools, and external loops, including agent workflows that inspect code and operate across a task environment (platform.claude.com). In practice, that makes the model good at classification. It can group similar strategies, spot repeated indicator logic, identify suspicious constructs, and flag scripts that deserve human review. A separate open-source TradingView automation project shows how an LLM can be wired into the Pine editor, charts, and Strategy Tester through tool bridges, letting it read chart state, write code, and run backtests inside TradingView itself (github.com). None of that gives the model market intuition. It gives it reach. And reach is useful when the real bottleneck is triage. If you hand a person 1,500 scalping scripts, they will spend days just opening files and recognizing patterns. If you hand them to a model, it can start building a taxonomy almost immediately: which scripts are moving-average crossovers in disguise, which ones depend on higher-timeframe requests, which ones rely on intrabar behavior, which ones are just parameter sweeps of the same core rule set. That is the hidden lesson in the video. The value is not that Claude can magically identify the winning bot from a heap of backtests. The value is that it can shrink the heap into something a trader can actually examine before risking money on a strategy that only ever worked in the past (youtube.com, tradingview.com).
