LLMs can become useful analysis assistants when they have structured data and realโtime market context, meaning they have enough relevant information for the task.
Getting specific indicators, news, and candles into an LLM is difficult, even with search tools integrated. You still need structured inputs, the right context, and a way to ensure everything the model receives is accurate and relevant.
For this reason, I built a system that prepares and delivers all required market data to LLMs in a clean, structured, and consistent format.
I'm excited to share my ๐๐ซ๐๐๐ข๐ง๐ ๐๐ข๐ ๐ง๐๐ฅ ๐๐๐ซ๐ฏ๐๐ซ, a researchโdriven system for developing and evaluating crypto signals with LLMs, quantitative models, and historical data.
๐๐ฒ๐ ๐ณ๐ฒ๐ฎ๐๐๐ฟ๐ฒ๐
โข ๐๐ถ๐ป๐ฎ๐ป๐ฐ๐ฒ ๐บ๐ฎ๐ฟ๐ธ๐ฒ๐ ๐ฑ๐ฎ๐๐ฎ including candles, orderโbook metrics, recent trades, and technical indicators
โข ๐ช๐ฒ๐ฏโ๐๐ฒ๐ฎ๐ฟ๐ฐ๐ต ๐ฐ๐ผ๐ป๐๐ฒ๐
๐ for crypto news, whale activity, policy, macro events, exchange updates, and whale alerts
โข ๐ข๐ฝ๐ฒ๐ป๐ฅ๐ผ๐๐๐ฒ๐ฟ ๐บ๐ผ๐ฑ๐ฒ๐น ๐๐ฒ๐น๐ฒ๐ฐ๐๐ถ๐ผ๐ป for choosing free or paid models from OpenAI, Google, Anthropic, Mistral, and others
โข ๐๐๐ ๐๐ผ๐๐ถ๐ป๐ด and repeated iterations for consistency checks
โข ๐ฃ๐ฟ๐ผ๐บ๐ฝ๐ ๐ณ๐ถ๐น๐ฒ๐ mapped to models and iteration cycles
โข ๐ค๐๐ฎ๐ป๐ ๐บ๐ผ๐ฑ๐ฒ๐น๐ for additional market evidence, with selected model families running in parallel
โข ๐๐ถ๐๐๐ผ๐ฟ๐ถ๐ฐ๐ฎ๐น ๐ฏ๐ฎ๐ฐ๐ธ๐๐ฒ๐๐๐ถ๐ป๐ด before live signal analysis
โข ๐ฃ๐ฒ๐ฟ๐ณ๐ผ๐ฟ๐บ๐ฎ๐ป๐ฐ๐ฒ ๐บ๐ฒ๐๐ฟ๐ถ๐ฐ๐ including win rate, returns, drawdown, profit factor, buyโandโhold comparison, outperformance, LLM agreement, direction accuracy, confidence calibration, and cost
โข ๐ง๐ฒ๐น๐ฒ๐ด๐ฟ๐ฎ๐บ ๐ฎ๐น๐ฒ๐ฟ๐๐ for qualifying signals
โข ๐ฆ๐ฎ๐๐ฒ๐ฑ ๐ฐ๐ผ๐ป๐ณ๐ถ๐ด๐๐ฟ๐ฎ๐๐ถ๐ผ๐ป๐ and prompt files for reproducible experiments
โข ๐ฆ๐ฒ๐น๐ณโ๐น๐ฎ๐ฏ๐ฒ๐น๐ถ๐ป๐ด ๐ ๐ ๐ฑ๐ฎ๐๐ฎ๐๐ฒ๐ signals are saved automatically, and the system later checks real future candles to label each one, giving you clean groundโtruth data for future agent training
โข ๐๐ผ๐ฐ๐ฎ๐น ๐ฑ๐ฎ๐๐ต๐ฏ๐ผ๐ฎ๐ฟ๐ฑ for live signal analysis, testing, configurations, prompts, settings, logs, and ML data
The platform is designed for continuous research, experimentation, and strategy improvement. It generates signals and alerts but does not place exchange orders automatically.
๐ Project: https://github.com/haidarali0/Trading-Signal-Server
Top comments (1)
Hey Haidar, this resonates a lot โ I've been building in a similar space (an AI DeFi agent using Coinbase AgentKit + x402 micropayments on Base, a Uniswap V4 trading agent with a GAN/teacher-student distillation pipeline, and a trade-copier EA), and the "structured data + real-time context is the actual bottleneck, not the LLM itself" problem is exactly what I kept running into too.
Your LLM voting + historical backtesting setup is the piece I haven't solved as cleanly โ I ended up going the distillation route (large teacher model โ fine-tuned small student) partly to sidestep exactly the consistency problem voting is trying to address at inference time. Have you compared voting-across-models against something like that, or found voting to be cheaper/more reliable in practice for signal generation specifically?
Also, on the order-book/candle side โ are you pulling those live per-signal, or snapshotting and caching on an interval? With Binance data specifically, I've hit issues where indicator values computed just before vs. just after a candle close diverge enough to flip a signal, and I'm curious whether your LLM voting step absorbs that noise or whether it's still sensitive to it.