Skip to content

RKiding/Awesome-finance-skills / skills/alphaear-predictor

alphaear-predictor

Unverified

Market prediction skill using Kronos. Use when user needs finance market time-series forecasting or news-aware finance market adjustments.

FileKindSize
SKILL.mdskill md2.1 KB
references/PROMPTS.mdmarkdown1.0 KB
scripts/__init__.pyscript0 B
scripts/forecast_agent.pyscript2.9 KB
scripts/json_utils.pyscript6.2 KB
scripts/kronos_predictor.pyscript7.6 KB
scripts/predictor/exports/models/kronos_news_v1_20260101_0015.ptother1.2 MB
scripts/predictor/model/__init__.pyscript394 B
scripts/predictor/model/kronos.pyscript30.4 KB
scripts/predictor/model/module.pyscript22.3 KB
scripts/prompts/fin_agent.pyscript6.6 KB
scripts/prompts/forecast_analyst.pyscript1.9 KB
scripts/prompts/intent_agent.pyscript2.2 KB
scripts/prompts/isq_prompt_generator.pyscript1.8 KB
scripts/prompts/report_agent.pyscript18.1 KB
scripts/prompts/trend_agent.pyscript7.4 KB
scripts/prompts/visualizer.pyscript2.0 KB
scripts/schema/isq_template.pyscript13.6 KB
scripts/schema/models.pyscript6.1 KB
scripts/utils/__init__.pyscript25 B
scripts/utils/database_manager.pyscript21.3 KB
scripts/utils/json_utils.pyscript6.2 KB
scripts/utils/llm/capability.pyscript2.8 KB
scripts/utils/llm/factory.pyscript3.6 KB
scripts/utils/llm/router.pyscript2.8 KB
scripts/utils/logging_setup.pyscript1.0 KB
scripts/utils/predictor/evaluation.pyscript5.6 KB
scripts/utils/predictor/kline_generate.pyscript7.6 KB
scripts/utils/predictor/model/__init__.pyscript394 B
scripts/utils/predictor/model/kronos.pyscript30.4 KB
scripts/utils/predictor/model/module.pyscript22.3 KB
scripts/utils/predictor/training.pyscript18.3 KB
scripts/utils/search_tools.pyscript28.6 KB
scripts/utils/stock_tools.pyscript9.8 KB
tests/test_predictor.pyscript1.0 KB

scripts/kronos_predictor.py

View on GitHub
import torch
import pandas as pd
import numpy as np
from datetime import datetime
from typing import List, Optional
from loguru import logger
from pandas.tseries.offsets import BusinessDay
from dotenv import load_dotenv

# Load environment variables
load_dotenv()

# Fix for Kronos internal imports
import sys
import os
KRONOS_DIR = os.path.join(os.path.dirname(__file__), 'predictor')
if KRONOS_DIR not in sys.path:
    sys.path.append(KRONOS_DIR)

import glob
from sentence_transformers import SentenceTransformer

from .predictor.model import Kronos, KronosTokenizer, KronosPredictor
from .schema.models import KLinePoint

class KronosPredictorUtility:
    """
    Kronos 时序预测工具类
    负责模型加载、推理以及数据结构转换
    """
    _instance = None
    _predictor = None

    def __new__(cls, *args, **kwargs):
        if not cls._instance:
            cls._instance = super(KronosPredictorUtility, cls).__new__(cls)
        return cls._instance

    def __init__(self, device: Optional[str] = None):
        if self._predictor is not None:
            return
            
        try:
            if not device:
                device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
            
            logger.info(f"🔮 Loading Kronos Model on {device}...")
            
            # 1. Load Embedder (SentenceTransformer)
            model_name = os.getenv('EMBEDDING_MODEL', 'sentence-transformers/all-MiniLM-L6-v2')  # Match training
            try:
                self.embedder = SentenceTransformer(model_name, device=device, local_files_only=True)
            except Exception:
                logger.warning(f"⚠️ Local embedder {model_name} not found. Downloading...")
                self.embedder = SentenceTransformer(model_name, device=device)

            # 2. Load Kronos Base
            try:
                tokenizer = KronosTokenizer.from_pretrained("NeoQuasar/Kronos-Tokenizer-base", local_files_only=True)
                model = Kronos.from_pretrained("NeoQuasar/Kronos-base", local_files_only=True)
            except Exception:
                logger.warning("⚠️ Local Kronos cache not found. Attempting to download...")
                tokenizer = KronosTokenizer.from_pretrained("NeoQuasar/Kronos-Tokenizer-base")
                model = Kronos.from_pretrained("NeoQuasar/Kronos-base")
            
            # 3. Load Trained News Projector Weights
            # Check predictor/exports/models directory
            models_dir = os.path.join(KRONOS_DIR, "exports/models")
            # Security: Only load files matching the expected prefix to reduce supply-chain risk
            model_files = glob.glob(os.path.join(models_dir, "kronos_news_*.pt"))
            
            if model_files:
                latest_model = max(model_files, key=os.path.getctime)
                logger.info(f"🔄 Loading trained news weights from {latest_model}...")
                try:
                    logger.info(f"🔄 Loading trained news weights from {latest_model}...")
                    # Security: Use weights_only=True to prevent pickle-based arbitrary code execution (PyTorch 1.13+)
                    try:
                        checkpoint = torch.load(latest_model, map_location=device, weights_only=True)
                    except (TypeError, RuntimeError, AttributeError):
                        # Fallback for older torch versions or complex objects if necessary
                        logger.warning(f"⚠️ Safe loading failed. Falling back to legacy load for {latest_model}")
                        checkpoint = torch.load(latest_model, map_location=device)
                    # The checkpoint contains 'news_proj_state_dict'
                    if 'news_proj_state_dict' in checkpoint:
                        if not hasattr(model, 'news_proj') or model.news_proj is None:
                            import torch.nn as nn
                            news_dim = checkpoint.get('news_dim', 384)
                            model.news_proj = nn.Linear(news_dim, model.d_model).to(device)
                        
                        model.news_proj.load_state_dict(checkpoint['news_proj_state_dict'])
                        logger.success("✅ News-Aware Projection Layer loaded!")
                        self.has_news_model = True
                    else:
                        logger.warning("⚠️ Checkpoint found but missing 'news_proj_state_dict'. Using base model.")
                        self.has_news_model = False
                except Exception as e:
                    logger.error(f"❌ Failed to load trained weights: {e}. Using base model.")
                    self.has_news_model = False
            else:
                logger.info("ℹ️ No trained news models found. Using base model.")
                self.has_news_model = False
            
            tokenizer = tokenizer.to(device)
            model = model.to(device)
            
            self._predictor = KronosPredictor(model, tokenizer, device=device, max_context=512)
            logger.info("✅ Kronos Model loaded successfully.")
        except Exception as e:
            logger.error(f"❌ Failed to load Kronos Model: {e}")
            self._predictor = None
            self.has_news_model = False

    def get_base_forecast(self, df: pd.DataFrame, lookback: int = 20, pred_len: int = 5, news_text: Optional[str] = None) -> List[KLinePoint]:
        """
        生成原始模型预测
        """
        if self._predictor is None:
            logger.error("Predictor not initialized.")
            return []

        if len(df) < lookback:
            logger.warning(f"Insufficient historical data ({len(df)}) for lookback ({lookback}).")
            return []

        # 获取最后 lookback 条数据
        x_df = df.iloc[-lookback:].copy()
        x_timestamp = pd.to_datetime(x_df['date']) # Ensure datetime
        last_date = x_timestamp.iloc[-1]
        
        # 生成未来时间戳
        future_dates = pd.date_range(start=last_date + BusinessDay(1), periods=pred_len, freq='B')
        y_timestamp = pd.Series(future_dates)

        # Embedding News if available
        news_emb = None
        if news_text and getattr(self, 'has_news_model', False) and hasattr(self, 'embedder'):
            try:
                # Truncate to avoid too long text
                emb = self.embedder.encode(news_text[:1000])
                news_emb = emb # KronosPredictor expects numpy array or tensor
            except Exception as e:
                logger.error(f"Failed to encode news: {e}")

        try:
            # 预测所需的列
            cols = ['open', 'high', 'low', 'close', 'volume']
            pred_df = self._predictor.predict(
                df=x_df[cols],
                x_timestamp=x_timestamp,
                y_timestamp=y_timestamp,
                pred_len=pred_len,
                T=1.0, 
                top_p=0.9, 
                sample_count=1,
                verbose=False,
                news_emb=news_emb
            )
            
            # 转换为 KLinePoint
            results = []
            for date, row in pred_df.iterrows():
                results.append(KLinePoint(
                    date=date.strftime("%Y-%m-%d"),
                    open=float(row['open']),
                    high=float(row['high']),
                    low=float(row['low']),
                    close=float(row['close']),
                    volume=float(row['volume'])
                ))
            return results
        except Exception as e:
            logger.error(f"Forecast generation failed: {e}")
            return []

# Singleton instance for easy access
# Usage: predictor = KronosPredictorUtility()
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 472kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall