正在打开每日关注
正在打开每日关注
解析Kronos模型族、Qlib数据处理、微调流程与开源边界。
Wall Street's worst nightmare isn't regulation. It's a GitHub repo. Kronos. An AI trained from scratch on 12 billion candles from 45 exchanges, and it reads charts as a native language. The numbers: - 93% better at ranking price forecasts than the best time-series model that exists - Zero-shot on any asset it's never seen. BTC, Nvidia, forex, whatever - 4M parameters at the small end, so it runs on a laptop - Peer reviewed at AAAI 2026 Banks survived 2008, the fines, the hearings... none of it mattered because the real moat was models nobody else had. This repo is inside the moat. MIT license, weights on Hugging Face, free. Star it before your fund manager does.
TradingView became 10x more valuable the day I connected it to Grok. That simple workflow helped me make over $17,129 while making better trading decisions. Here are the 8 SECRET prompts behind it: (BOOKMARK BEFORE I DELETE THIS)🧵 https://t.co/g95sevRZFo
是什么
Kronos 是 GitHub 上 shiyu-coder/Kronos 的金融时间序列模型项目。README 截图显示它包含 Model Zoo、MIT license、Hugging Face Hub 权重链接，以及安装、预测和微调说明。推文所称 120 亿根 K 线、45 家交易所、93% 提升和 AAAI 2026，截图无法独立证实。
架构与核心设计
其公开流程是 tokenizer + predictor 两阶段：先从 Qlib 目录运行 finetune/qlib_data_preprocess.py，生成 train/validation/test pickle；再微调 tokenizer 与 predictor，脚本支持 torchrun 多 GPU。Model Zoo 按容量列出 mini/small/base/large，参数量 4.1M 至 499.2M、上下文 512 或 2048；large 行的 open-source 标记为红叉。
值得借鉴
把模型规模、上下文和开源状态表格化，便于按算力选型；同时显式配置数据集、检查点、回测输出和预训练路径，形成从数据准备到评估的可复现路径。KronosPredictor 与安装命令降低了上手门槛。
边界与风险 金融回测容易受数据泄漏、时间切分、交易成本和市场变化影响。“零样本”“战胜最佳模型”以及引用推文中的获利叙述应视为待验证宣传，使用前需核对论文、权重、数据许可和基准脚本，不能将预测直接当作投资建议。