TimesFM：Google Research 时间序列基础模型完全指南

model.compile(
    timesfm.ForecastConfig(
        max_context=1024,
        max_horizon=256,
        use_continuous_quantile_head=True,  # 启用连续分位数
        # 可选 30M 分位数头
    )
)

# 输出形状: (batch_size, horizon, num_quantiles)
# num_quantiles 包括均值 + 10个分位数 (10th-90th)

import torch
import numpy as np
import timesfm

# 设置高精度矩阵乘法
torch.set_float32_matmul_precision("high")

# 加载模型
model = timesfm.TimesFM_2p5_200M_torch.from_pretrained(
    "google/timesfm-2.5-200m-pytorch"
)

# 编译模型（加速推理）
model.compile(
    timesfm.ForecastConfig(
        max_context=1024,
        max_horizon=256,
        normalize_inputs=True,
        use_continuous_quantile_head=True,
        force_flip_invariance=True,
        infer_is_positive=True,
        fix_quantile_crossing=True,
    )
)

# 预测
point_forecast, quantile_forecast = model.forecast(
    horizon=12,  # 预测未来 12 个时间点
    inputs=[
        np.linspace(0, 1, 100),        # 输入序列 1
        np.sin(np.linspace(0, 20, 67)), # 输入序列 2
    ],
)

print(f"点预测形状: {point_forecast.shape}")      # (2, 12)
print(f"分位数预测形状: {quantile_forecast.shape}") # (2, 12, 10)

# point_forecast: 每个序列的点预测
# shape: (num_series, horizon)

# quantile_forecast: 分位数预测
# shape: (num_series, horizon, num_quantiles)
# quantiles: [mean, 10th, 20th, ..., 90th]

# 提取特定分位数
mean_pred = quantile_forecast[:, :, 0]      # 均值
lower_80 = quantile_forecast[:, :, 1]       # 10th 分位数（下界）
upper_80 = quantile_forecast[:, :, -1]      # 90th 分位数（上界）

# 同时预测多个序列
inputs = [
    np.sin(np.linspace(0, 10, 500)),   # 序列 1: 正弦波
    np.random.randn(500),               # 序列 2: 随机
    [1, 2, 3, 4, 5, 6, 7, 8, 9, 10], # 序列 3: 线性增长
]

point_pred, quantile_pred = model.forecast(
    horizon=24,
    inputs=inputs,
)

# 输出形状: (3, 24) 和 (3, 24, 10)

# 带有外部变量的预测
holidays = [0, 0, 0, 1, 0, 0, 0] * 4  # 节假日标记
promotions = [0, 1, 0, 0, 0, 0, 1] * 4  # 促销标记

covariates = np.stack([holidays, promotions], axis=0)

point_pred, quantile_pred = forecast_with_covariates(
    model=model,
    covariates=covariates,
    horizon=24,
    context=input_series,
)

<!-- AGENTS.md 内容示例 -->
# TimesFM AI Agent Skills

## Forecasting Agent
Use TimesFM to forecast time series data.

## Anomaly Detection
Detect anomalies using quantile predictions.

## Usage
1. Load your time series data
2. Call TimesFM forecast
3. Interpret quantile predictions

class PatchFormer(nn.Module):
    """TimesFM 核心架构"""
    
    def __init__(self, config):
        super().__init__()
        # 补丁嵌入
        self.patch_embed = nn.Linear(
            config.patch_len,
            config.d_model
        )
        
        # Transformer 解码器
        self.blocks = nn.ModuleList([
            TransformerBlock(config)
            for _ in range(config.n_layers)
        ])
        
        # 输出头
        self.head = nn.Linear(
            config.d_model,
            config.patch_len
        )