实现网
177人浏览/1人投稿
357天前
待托管赏金
中文版本
要求: 博士学历程序员，精通Python、机器学习、神经网络、PyTorch、TensorFlow、多层感知机、Transformers和时间序列分析。需具备基础英语能力，能够有效沟通并理解所提供的英文技术研究文档。
预算: 1,500元人民币
时间: 5天
提供数据: 您将使用我提供的四个特定时间序列数据集：一个货币对、一个股票价格、中国采购经理指数(PMI)和德国电力价格。
第一阶段：实时预测实现
本项目涉及扩展一个已经使用Nixtla的NeuralForecast框架成功测试的预训练TimeMixer神经网络模型。该模型目前在金融时间序列数据上执行准确的30天预测，并已证明有效果。现有代码包含季节性嵌入功能和在测试环境中表现出色的参数。您的任务是扩展这个工作模型，生成超出最后可用数据点的真实预测，为每一天创建逐步的日度预期收益和价格预测，直到第30天。您将实现Optuna优化来为每个金融工具微调参数，并创建一个可以处理任何上传数据集的动态管道。系统必须保持与测试环境中展示的相同准确度水平，预测输出必须与提供的参考格式完全匹配。
第二阶段：投注规模算法实现
完成实时预测后，您将基于我将提供的研究文档实现一个综合投注规模系统。该文档已包含代码片段，因此不需要进行研究工作。这涉及创建一个元标记框架，用于确定最优投注规模而非仅仅是市场方向，使用从TimeMixer模型预测中得出的基于概率的规模计算。实现包括带有止盈、止损和时间退出的三重障碍风险管理方法，以及动态头寸规模计算。最终系统必须完全动态化，允许用户上传任何金融工具数据集并自动执行所有预测和投注规模步骤，以提供清晰的资本配置建议。
English Version
Requirements: PhD-level programmer proficient in Python, Machine Learning, Neural Networks, PyTorch, TensorFlow, Multi-layer Perceptrons, Transformers, and Time Series Analysis. Basic English proficiency required for effective communication and comprehension of technical research documentation provided in English.
Budget: 1,500 Yuan
Timeline: 5 days
Data Provided: You will work with four specific time series datasets that I will provide: one currency pair, one stock price, China's Purchasing Managers Index (PMI), and Germany's electricity prices.
Phase 1: Real-Time Forecasting Implementation
This project involves extending a pre-trained TimeMixer neural network model that has already been successfully tested using Nixtla's NeuralForecast framework. The model currently performs accurate 30-day forecasting on financial time series data with proven results. The existing code includes seasonal embedding functionality and parameters that have demonstrated strong performance in testing environments. Your task is to extend this working model to generate real predictions beyond the last available data point, creating step-by-step daily forecasts for expected returns and prices for each day up to the 30th day. You will implement Optuna optimization to fine-tune parameters for each financial instrument and create a dynamic pipeline that can process any uploaded dataset. The system must maintain the same level of accuracy demonstrated in the tested environment, and the forecast output must match the provided reference format exactly.
Phase 2: Bet Sizing Algorithm Implementation
After completing the real-time forecasting, you will implement a comprehensive bet sizing system based on the research documentation that I will provide. The documentation already has the code snippets, therefore research will not be needed. This involves creating a meta-labeling framework that determines optimal bet sizes rather than just market direction, using probability-based sizing derived from the TimeMixer model's predictions. The implementation includes triple-barrier risk management methods with profit-taking, stop-loss, and time-based exits, along with dynamic position sizing calculations. The final system must be fully dynamic, allowing users to upload any financial instrument dataset and automatically follow all forecasting and bet sizing steps to provide clear capital allocation recommendations.
