导师说:别再死磕LSTM了!20行Python代码,直接搞懂Prophet时间序列预测

import warningsimport matplotlib.pyplot as pltimport numpy as npimport pandas as pdfrom prophet import Prophetimport seaborn as snsimport yfinance as yf# ==========================================# 1. 从雅虎财经获取数据 (以 AAPL 为例)# ==========================================ticker = "AAPL"# 可替换为 TSLA, NVDA 等股票代码print(f"正在从雅虎财经下载 {ticker} 历史价格数据...")# 获取过去 3 年的数据data = yf.download(ticker, start="2010-01-01", end="2026-08-01")# 处理多层索引 (yfinance 近期更新可能返回 MultiIndex 列)if isinstance(data.columns, pd.MultiIndex):  data = data["Close"]else:  data = data[["Close"]]# Prophet 严格要求两列数据: 'ds' (日期) 和 'y' (目标数值)df = data.reset_index()df.columns = ["ds", "y"]# 移除时区信息（Prophet 处理无时区时间序列稳定性更好）df["ds"] = pd.to_datetime(df["ds"]).dt.tz_localize(None)print(f"成功获取 {len(df)} 条交易日数据。")# ==========================================# 2. 构建并训练 Prophet 模型# ==========================================# 股票市场数据包含明显的周季节性（交易日）和年季节性model = Prophet(    daily_seasonality=False,    weekly_seasonality=True,    yearly_seasonality=True,    changepoint_prior_scale=0.05,  # 变点灵活度控制，防止过拟合    seasonality_prior_scale=10.0,  # 季节性先验强度    interval_width=0.95,  # 设置 95% 置信区间)# 拟合模型model.fit(df)# 创建未来 180 天的预测时间序列（剔除周末以符合交易日特点）future = model.make_future_dataframe(periods=180, freq="D")future = future[future["ds"].dt.dayofweek < 5]  # 保留周一至周五# 执行预测forecast = model.predict(future)# ==========================================# 3. 绘制 4 个精美画板图表# ==========================================fig = plt.figure(figsize=(16, 12), dpi=120)# ------------------------------------------# 图 1：历史收盘价与未来 Prophet 预测走势图# ------------------------------------------ax1 = plt.subplot(2, 2, 1)# 真实观测值ax1.scatter(    df["ds"],    df["y"],    color="#1f77b4",    s=12,    alpha=0.6,    label="Historical Close Price",)# 预测趋势线ax1.plot(    forecast["ds"], forecast["yhat"], color="#d62728", lw=2, label="Forecast")# 95% 置信区间填充ax1.fill_between(    forecast["ds"],    forecast["yhat_lower"],    forecast["yhat_upper"],    color="#d62728",    alpha=0.15,    label="95% Confidence Interval",)ax1.set_title(    f"1. {ticker} Stock Price Prediction (Prophet)",    fontsize=12,    fontweight="bold",    pad=10,)ax1.set_ylabel("Price ($)")ax1.legend(loc="upper left", frameon=True)# ------------------------------------------# 图 2：趋势解构与变点（Changepoints）分析图# ------------------------------------------ax2 = plt.subplot(2, 2, 2)# 整体趋势线ax2.plot(    forecast["ds"],    forecast["trend"],    color="#2ca02c",    lw=2.5,    label="Overall Trend",)# 标注 Prophet 自动探测的主要变点（取变化显著前10个）changepoints = model.changepointsif len(changepoints) > 0:  deltas = np.abs(model.params["delta"].squeeze())  top_cps = changepoints.iloc[np.argsort(deltas)[-8:]]for cp in top_cps:    ax2.axvline(        x=cp, color="#ff7f0e", linestyle="--", alpha=0.6, lw=1.2, label="_nolegend_"    )  ax2.axvline(      x=top_cps.iloc[0],      color="#ff7f0e",      linestyle="--",      alpha=0.6,      lw=1.2,      label="Significant Changepoints",  )ax2.set_title("2. Underlying Trend & Detected Changepoints",    fontsize=12,    fontweight="bold",    pad=10,)ax2.set_ylabel("Trend Magnitude")ax2.legend(loc="upper left", frameon=True)# ------------------------------------------# 图 3：每周与每年的季节性模式 (Seasonality)# ------------------------------------------ax3 = plt.subplot(2, 2, 3)# 计算一周内各天的平均影响weekly_df = forecast.groupby(forecast["ds"].dt.day_name())["weekly"].mean()days_order = ["Monday","Tuesday","Wednesday","Thursday","Friday",]weekly_df = weekly_df.reindex(days_order)# 柱状图展示周内效应colors = ["#4c72b0"if v >= 0 else"#c44e52"for v in weekly_df.values]ax3.bar(    weekly_df.index,    weekly_df.values,    color=colors,    alpha=0.85,    width=0.5,    edgecolor="black",    lw=0.5,)ax3.axhline(0, color="gray", linewidth=0.8)ax3.set_title("3. Weekly Seasonality Pattern (Day of Week Effect)",    fontsize=12,    fontweight="bold",    pad=10,)ax3.set_ylabel("Price Impact ($)")# ------------------------------------------# 图 4：模型训练残差（Errors）分布诊断图# ------------------------------------------ax4 = plt.subplot(2, 2, 4)# 合并真实值与拟合值，计算残差merged = pd.merge(df, forecast[["ds", "yhat"]], on="ds")residuals = merged["y"] - merged["yhat"]# 绘制残差分布核密度图sns.histplot(    residuals,    kde=True,    ax=ax4,    color="#8c564b",    bins=30,stat="density",    alpha=0.4,)ax4.axvline(0, color="red", linestyle="--", linewidth=1.5, label="Zero Error")ax4.set_title("4. Model Residuals Distribution (In-sample Error)",    fontsize=12,    fontweight="bold",    pad=10,)ax4.set_xlabel("Residual (Actual - Predicted)")ax4.set_ylabel("Density")ax4.legend(loc="upper right")# 调整子图间距并展示plt.tight_layout()plt.show()