关于「前期拟合是入场券:均方根预测误差、凸包与留出检验」,《合成控制与单站改版》给出的定量结论是什么?

结论:在第零号面板上,合成对照的拟合期误差是 6.39,留出期误差 4.00。自由回归的拟合期误差只有 3.11,留出期却是 15.51。两百个面板的平均更清楚:合成对照拟合 6.57、留出 9.29,留出与拟合之比 1.41;自由回归拟合 2.96、留出 19.72,比值 6.66。捐赠站简单平均的拟合误差 18.09,留出 19.55;平移后的平均拟合 13.63,留出 15.23;最好的单个站拟合 13.82,留出 16.39。后三者的比值都接近一,因为它们几乎没有在拟合期「学」到什么,自然也不存在过拟合,但它们的绝对误差都比合成对照大得多。

数据口径:在一个三十四站的示意面板上,真实提升设为每周十二单,前后对比报出 37.77,主要误差来自季节;用单纯形约束权重把三十三个未改版的捐赠站加权成合成对照站,估计为 8.83。 两百个面板上,前后对比的均方根误差为 35.61,合成对照为 5.42。

English version

对照数据表

下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: 1.41, 13.63, 13.82, 15.23, 15.51, 16.39

panel 0 method fit wk 0-39 check wk 40-51
simplex6.39 4.00
free OLS3.11 15.51
donor mean11.52 13.72
mean+shift11.50 14.24
best single13.63 19.30
200 panels, means:
simplex6.57 9.29 ratio 1.41
free OLS2.96 19.72 ratio 6.66
donor mean18.09 19.55 ratio 1.08
mean+shift13.63 15.23 ratio 1.12
best single13.82 16.39 ratio 1.19
查看完整验算程序
# Chapter 3. p33: hold out the last 12 pre weeks before believing a fit.
# Panel as in p11 (illustrative). Weights are fitted on weeks 0-39
# only and judged on weeks 40-51, still before the redesign. Methods:
# synthetic control (simplex), free OLS, the plain donor mean, the
# donor mean shifted to the site's level (a DiD-style comparison), and
# the single best donor. One panel, then means over 200 panels.
import math
import random
from operator import mul

# 示意面板:每站周订单 = 基准 + 季节载荷 * 季节 + 大盘载荷 * 大盘 + 噪声
T0, POST, NS = 52, 12, 34


def panel(rep=0, n=NS, t0=T0, post=POST, tau=12.0, start=10):
    rng = random.Random(20261113 + 7919 * rep)
    weeks = t0 + post
    mk = [0.0]
    for _ in range(weeks - 1):
        mk.append(mk[-1] + rng.gauss(0.05, 0.25))
    ss = [math.sin(2.0 * math.pi * (start + t) / 52.0)
          for t in range(weeks)]
    ys, cf, pars = [], [], []
    for i in range(n):
        if i == 0:
            mu, ra, rb, sd = 240.0, 0.20, 1.30, 7.0
        else:
            mu = rng.uniform(120.0, 360.0)
            ra = rng.uniform(0.04, 0.22)
            rb = rng.uniform(0.5, 1.7)
            sd = rng.uniform(5.0, 9.0)
        a, b = mu * ra, mu * rb / 20.0
        y = [mu + a * ss[t] + b * mk[t] + rng.gauss(0.0, sd)
             for t in range(weeks)]
        cf.append(y[:])
        if i == 0:
            for t in range(t0, weeks):
                y[t] = y[t] + tau
        ys.append(y)
        pars.append((mu, a, b, sd))
    return ys, cf, pars, ss, mk


def dot(a, b):
    return math.fsum(map(mul, a, b))


def mean(xs):
    return math.fsum(xs) / len(xs)


def rmspe(xs):
    return math.sqrt(math.fsum([x * x for x in xs]) / len(xs))


def gram(ys, t0):
    n = len(ys)
    pre = [y[:t0] for y in ys]
    g = [[0.0] * n for _ in range(n)]
    for i in range(n):
        for k in range(i, n):
            v = dot(pre[i], pre[k]) / t0
            g[i][k] = v
            g[k][i] = v
    return g


def solve(a, b):
    m = len(b)
    r = [row[:] + [b[i]] for i, row in enumerate(a)]
    for c in range(m):
        p = max(range(c, m), key=lambda i: abs(r[i][c]))
        r[c], r[p] = r[p], r[c]
        for i in range(c + 1, m):
            f = r[i][c] / r[c][c]
            if f:
                r[i] = [x - f * y for x, y in zip(r[i], r[c])]
    x = [0.0] * m
    for i in range(m - 1, -1, -1):
        s = math.fsum([r[i][k] * x[k] for k in range(i + 1, m)])
        x[i] = (r[i][m] - s) / r[i][i]
    return x


def eq_qp(g, c, sup):
    m = len(sup)
    a = [[2.0 * g[i][k] for k in sup] + [1.0] for i in sup]
    a.append([1.0] * m + [0.0])
    b = [2.0 * c[i] for i in sup] + [1.0]
    return solve(a, b)[:m]


def simplex_ls(g, c, tol=1e-10):
    # 精确有效集法:min w'Gw - 2c'w,w >= 0,权重和为 1
    j = len(c)
    k0 = min(range(j), key=lambda k: g[k][k] - 2.0 * c[k])
    w = [0.0] * j
    w[k0] = 1.0
    sup = [k0]
    for _ in range(4 * j + 20):
        gr = [2.0 * (math.fsum([g[i][k] * w[k] for k in sup]) - c[i])
              for i in range(j)]
        nu = -math.fsum([gr[k] for k in sup]) / len(sup)
        out = [k for k in range(j) if k not in sup]
        if not out:
            break
        kb = min(out, key=lambda k: gr[k] + nu)
        if gr[kb] + nu >= -tol:
            break
        sup = sorted(sup + [kb])
        while True:
            z = eq_qp(g, c, sup)
            if min(z) > 0.0:
                for k, v in zip(sup, z):
                    w[k] = v
                break
            step = min(w[k] / (w[k] - v) for k, v in zip(sup, z)
                       if v <= 0.0)
            for k, v in zip(sup, z):
                w[k] = w[k] + step * (v - w[k])
            sup = [k for k in sup if w[k] > 1e-12]
            for k in range(j):
                if k not in sup:
                    w[k] = 0.0
    return w


def synth(ys, g, tr, donors):
    gs = [[g[i][k] for k in donors] for i in donors]
    cs = [g[i][tr] for i in donors]
    w = simplex_ls(gs, cs)
    weeks = len(ys[tr])
    gap = [ys[tr][t] - dot(w, [ys[d][t] for d in donors])
           for t in range(weeks)]
    return w, gap


def ratio(gap, t0=T0):
    return rmspe(gap[t0:]) / rmspe(gap[:t0])


def placebo(ys, g, tr, pool, t0=T0):
    # 逐个捐赠站当处理站;处理站本身不进入安慰剂的捐赠池
    out = {}
    for j in pool:
        dn = [k for k in pool if k != j]
        out[j] = synth(ys, g, j, dn)[1]
    return out


def predict(ys, method, donors, t_fit):
    g = gram(ys, t_fit)
    gs = [[g[i][k] for k in donors] for i in donors]
    cs = [g[i][0] for i in donors]
    weeks = len(ys[0])
    if method == "simplex":
        w = simplex_ls(gs, cs)
    elif method == "free OLS":
        w = solve(gs, cs)
    elif method == "best single":
        k = min(range(len(donors)),
                key=lambda k: gs[k][k] - 2.0 * cs[k])
        w = [0.0] * len(donors)
        w[k] = 1.0
    else:
        w = [1.0 / len(donors)] * len(donors)
    yh = [dot(w, [ys[d][t] for d in donors]) for t in range(weeks)]
    if method == "mean+shift":
        sh = mean([ys[0][t] - yh[t] for t in range(t_fit)])
        yh = [v + sh for v in yh]
    return yh


if __name__ == "__main__":
    donors = list(range(1, NS))
    meths = ("simplex", "free OLS", "donor mean", "mean+shift",
             "best single")
    acc = {m: ([], []) for m in meths}
    for r in range(200):
        ys, cf, pars, ss, mk = panel(r)
        for m in meths:
            yh = predict(ys, m, donors, 40)
            fit = rmspe([ys[0][t] - yh[t] for t in range(40)])
            val = rmspe([ys[0][t] - yh[t] for t in range(40, T0)])
            acc[m][0].append(fit)
            acc[m][1].append(val)
            if r == 0 and m == meths[0]:
                print("panel 0  method        fit wk 0-39  check wk 40-51")
            if r == 0:
                print("         {:<12s} {:10.2f}  {:13.2f}".format(
                    m, fit, val))
    print("200 panels, means:")
    for m in meths:
        f, v = acc[m]
        print("         {:<12s} {:10.2f}  {:13.2f}   ratio {:.2f}".format(
            m, mean(f), mean(v), mean(v) / mean(f)))

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《合成控制法与单站改版因果、捐赠站加权重建反事实与安慰剂置换检验母典》第 3 章「前期拟合是入场券:均方根预测误差、凸包与留出检验」(版本 1.0.0). 5D Library. DOI 待发布

DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-304 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026304,
  author    = {Fan, Yafu},
  title     = {Synthetic Control and the Causal Effect of a Single-Site Redesign: Rebuilding the Counterfactual from Weighted Donor Sites and Placebo Permutation Tests},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论