5D Library › 做市商伽马敞口、零日到期期权与波动率收割动力学 · English

第二章 对冲反馈环与行权价钉死

一个方程

设做市商的净伽马库存为每单位价格 GG 股:价格变动 dSdS 时,保持对冲意味着交易 −GdS-G\,dS 股。设每交易一股使价格移动 λ\lambda。把两者的乘积写成一个无量纲数:

κ=λG。\kappa = \lambda G 。

κ>0\kappa > 0 表示做市商多头伽马:为保持对冲,他涨时卖、跌时买。κ<0\kappa < 0 表示空头伽马:涨时买、跌时卖。整章都来自这笔交易反馈回价格时发生的事。

推导

即时对冲。 若对冲与它所回应的价格变动同时成交,观察到的变动等于基本面变动 dS0dS_0 加上对冲本身的冲击: dS=dS0−κdS⇒dS=dS01+κ.dS = dS_0 – \kappa\, dS \quad\Longrightarrow\quad dS = \frac{dS_0}{1+\kappa}. 已实现波动率被乘以 1/(1+κ)1/(1+\kappa):做市商多头伽马时小于一,空头时大于一。κ=−1\kappa = -1 时方程没有有限解——价格必须跳空。

滞后对冲。 若上一次变动的对冲在此刻成交,dSt=dS0,t−κdSt−1dS_t = dS_{0,t} – \kappa\, dS_{t-1},这是一个一阶自回归过程,滞后一期自相关为 −κ-\kappa,单步方差为 σ2/(1−κ2)\sigma^2/(1-\kappa^2)。在很多步上加总,增量合计为 ∑dS0/(1+κ)\sum dS_0/(1+\kappa),所以在长时段上波动率同样被乘以 1/(1+κ)1/(1+\kappa)。

这正是关于动态对冲反馈效应的文献在连续时间里得到的结构 (Frey and Stremme 1997; Platen and Schweizer 1998):持有期权伽马多头的对冲者降低标的波动率,空头的对冲者提高它;当对冲的市场冲击接近市场深度时,放大倍数发散。

一个模拟,以及关于时间尺度的一处微妙之处

推导

# Chapter 2. Delta-hedging feedback. A dealer with gamma inventory G must
# trade -G x (price change) shares to stay hedged; each share traded moves the
# price by lam. Write kappa = lam x G (dimensionless). Two timings are
# compared.
#   Instant hedging:  dS = dS0 - kappa dS     ->  dS = dS0 / (1 + kappa)
#   Lagged hedging:   dS_t = dS0_t - kappa dS_{t-1}   (hedge of the last move
#   executes now)
# kappa > 0: dealer LONG gamma (sells rallies, buys dips); kappa < 0: SHORT
# gamma. 390 one-minute steps a day, 2,000 simulated days per kappa, fixed
# seed.
import math
import random

# fundamental one-minute volatility
SIG_MIN = 0.20 / math.sqrt(252 * 390)
N_MIN, N_DAYS = 390, 2000

def simulate(kappa, lagged, rng):
    minute_var, daily = 0.0, []
    ac_num = ac_den = 0.0
    for _ in range(N_DAYS):
        prev, tot = 0.0, 0.0
        for _ in range(N_MIN):
            e = rng.gauss(0.0, SIG_MIN)
            d = e - kappa * prev if lagged else e / (1.0 + kappa)
            minute_var += d * d
            ac_num += d * prev
            ac_den += prev * prev
            tot += d
            prev = d
        daily.append(tot)
    mvar = minute_var / (N_DAYS * N_MIN)
    dvar = sum(x * x for x in daily) / N_DAYS
    return (
        math.sqrt(mvar) / SIG_MIN,
        math.sqrt(dvar / N_MIN) / SIG_MIN,
        ac_num / ac_den,
    )

if __name__ == "__main__":
    rng = random.Random(56003)
    print("A. Instant hedging: realised vol relative to fundamental vol")
    print("   kappa    simulated   theory 1/(1+kappa)")
    for k in (0.5, 0.2, 0.0, -0.2, -0.5, -0.8):
        m, d, ac = simulate(k, False, rng)
        print("   {:+5.1f}   {:9.3f}   {:9.3f}".format(k, d, 1 / (1 + k)))
    print(
        "   kappa <= -1: no finite solution; the price must jump (liquidity"
        " gap)"
    )

    print(
        "\nB. Lagged hedging: one-minute vol, daily vol, lag-1"
        " autocorrelation"
    )
    print("   kappa   1-min vol   theory   daily vol   theory    autocorr")
    for k in (0.5, 0.2, 0.0, -0.2, -0.5, -0.8):
        m, d, ac = simulate(k, True, rng)
        print(
            "   {:+5.1f}   {:8.3f}  {:7.3f}   {:8.3f}  {:7.3f}   {:+8.3f}"
            .format(k, m, 1 / math.sqrt(1 - k * k), d, 1 / (1 + k), ac)
        )
    print(
        "\nReading: long gamma (kappa > 0) cuts DAILY volatility and makes"
        " minute"
    )
    print(
        "returns mean-revert; short gamma raises daily volatility and makes"
        " them"
    )
    print(
        "trend. With lagged hedging, one-minute volatility rises in BOTH"
        " cases, so"
    )
    print(
        "the dampening of long gamma shows up at the daily horizon, not tick"
        " by tick."
    )
# A. Instant hedging: realised vol relative to fundamental vol
#    kappa    simulated   theory 1/(1+kappa)
#     +0.5       0.641       0.667
#     +0.2       0.816       0.833
#     +0.0       1.004       1.000
#     -0.2       1.218       1.250
#     -0.5       1.987       2.000
#     -0.8       5.240       5.000
#    kappa <= -1: no finite solution; the price must jump (liquidity gap)
#
# B. Lagged hedging: one-minute vol, daily vol, lag-1 autocorrelation
#    kappa   1-min vol   theory   daily vol   theory    autocorr
#     +0.5      1.154    1.155      0.684    0.667     -0.500
#     +0.2      1.021    1.021      0.851    0.833     -0.197
#     +0.0      1.001    1.000      0.984    1.000     +0.001
#     -0.2      1.021    1.021      1.251    1.250     +0.198
#     -0.5      1.154    1.155      1.958    2.000     +0.499
#     -0.8      1.664    1.667      4.937    5.000     +0.800
#
# Reading: long gamma (kappa > 0) cuts DAILY volatility and makes minute
# returns mean-revert; short gamma raises daily volatility and makes them
# trend. With lagged hedging, one-minute volatility rises in BOTH cases, so
# the dampening of long gamma shows up at the daily horizon, not tick by tick.

A 段与即时对冲公式吻合:κ=+0.5\kappa = +0.5 时已实现波动率为基本面值的 0.641,理论值 0.667;κ=−0.8\kappa = -0.8 时为 5.240,理论值 5.000。极端处吻合较松,因为一个如此接近稳定边界的过程,其单日加总本身就很嘈杂。

B 段值得停下来看。滞后对冲下,多头伽马使一分钟收益均值回复(κ=+0.5\kappa = +0.5 时自相关 −0.500),空头伽马使其趋势化(κ=−0.5\kappa = -0.5 时 +0.499)——但一分钟波动率在两种情况下都上升,κ=±0.5\kappa = \pm 0.5 时为基本面值的 1.154 倍。熟悉的不对称只在日度时段出现:多头伽马 0.684,空头 1.958。

所以「多头伽马压低波动率」是一个关于测量时段的陈述。逐笔看,一个多头伽马的市场可能比完全没有对冲者的市场更忙乱;按日看,它更平静。任何对这一机制的实证检验都必须说明用的是哪个时段。

左:日度波动率相对基本面波动率之比 1/(1+\kappa),以及滞后对冲下的一分钟波动率 1/\sqrt{1-\kappa^2},随 \kappa 变化。右:滞后对冲下收益的滞后一期自相关 -\kappa:多头伽马为均值回复,空头伽马为动量。

证据

动量那一半的证据。 Baltussen、Da、Lammers 与 Martens 用指数数据发现,期权做市商与杠杆型交易所交易基金的对冲需求,能预测当日最后半小时沿当日早先方向的收益——即日内动量——且对冲需求越大效应越强 (Baltussen et al. 2021)。这正是 B 段中空头伽马、滞后对冲的预测,被测到了。在个股层面,Ni、Pearson、Poteshman 与 White 发现期权做市商的净伽马与标的股票的波动率相关,方向与对冲论证的预测一致 (Ni et al. 2021)。

行权价钉死

临近到期时,集中在某一行权价的未平仓量,恰好在该行权价处造出一个伽马尖峰(第一章)。Avellaneda 与 Lipkin 证明,当对冲者持有这份伽马的多头、且其交易有价格冲击时,股价会随到期临近被吸向该行权价 (Avellaneda and Lipkin 2003)。他们的机制有两部分,都已在第一章出现:伽马部分(高于行权价时卖、低于时买)与 charm 部分(Delta 随时钟衰减向 0 或 1)。

第四段程序直接实现这一机制。做市商持有行权价 100 的 QQ 张认购并保持 −QΔ-Q\Delta 股的对冲;每交易一股都移动价格,所以成交价满足不动点方程 S=F−η[Δ(S,t)−Δ(S0,0)]S = F – \eta\,[\Delta(S,t) – \Delta(S_0,0)],其中 FF 是不受扰动的价格,η\eta 是冲击乘以头寸。每个模拟交易日的开盘价在行权价上下一点之内均匀抽取。

推导

# Chapter 2. Strike pinning on expiry day (an Avellaneda-Lipkin type
# mechanism). Dealers hold Q calls struck at K = 100 expiring at the 16:00
# close and keep the hedge -Q x delta in shares. Each share they trade moves
# the price, so the traded price S solves   S = F - eta x (delta(S, t) -
# delta(S0, 0)),   eta = (impact per share) x Q. F is the price the market
# would have shown without the hedge (a random walk). eta > 0: dealers LONG
# the calls (long gamma). eta < 0: dealers SHORT them. Each simulated day
# starts uniformly within +/- 1.00 of the strike; 5-minute steps; the close is
# read at 15:55. "Pinned" means |S_close - K| < 0.10 (0.1 % of price).
import math
import random

K, SIG = 100.0, 0.20
YEAR_MIN = 252 * 390.0
STEPS = 77                    # 5-minute steps from 09:35 to 15:55
DT = 5.0 / YEAR_MIN

def Phi(x):
    return 0.5 * (1.0 + math.erf(x / math.sqrt(2.0)))

def delta(S, T):
    return Phi((math.log(S / K) + 0.5 * SIG * SIG * T) / (SIG * math.sqrt(T)))

def gam(S, T):
    d1 = (math.log(S / K) + 0.5 * SIG * SIG * T) / (SIG * math.sqrt(T))
    return math.exp(-0.5 * d1 * d1) / (
        math.sqrt(2 * math.pi) * S * SIG * math.sqrt(T)
    )

def solve(F, eta, d0, T, s_prev):
    g = lambda s: s - F + eta * (delta(s, T) - d0)
    s = s_prev
    # Newton from the previous price
    for _ in range(30):
        slope = 1.0 + eta * gam(s, T)
        if slope <= 0.05:
            break
        step = g(s) / slope
        s -= step
        if abs(step) < 1e-9:
            return s, False
    # branch lost (only possible for eta < 0): scan for all roots, take the
    # nearest
    grid = [F - 3.0 + 0.01 * i for i in range(601)]
    roots, prev = [], g(grid[0])
    for a, b in zip(grid, grid[1:]):
        gb = g(b)
        if prev == 0 or prev * gb < 0:
            lo, hi, glo = a, b, prev
            for _ in range(40):
                mid = 0.5 * (lo + hi)
                gm = g(mid)
                if (gm < 0) == (glo < 0):
                    lo, glo = mid, gm
                else:
                    hi = mid
            roots.append(0.5 * (lo + hi))
        prev = gb
    return min(roots, key=lambda r: abs(r - s_prev)), True

def run(eta, n, rng):
    pinned, dist, jumps = 0, 0.0, 0
    for _ in range(n):
        S0 = K + rng.uniform(-1.0, 1.0)
        T0 = (STEPS + 1) * DT
        d0 = delta(S0, T0)
        F, S = S0, S0
        for i in range(STEPS):
            F += S0 * SIG * math.sqrt(DT) * rng.gauss(0.0, 1.0)
            T = (STEPS - i) * DT
            S_new, lost = solve(F, eta, d0, T, S)
            if lost:
                jumps += 1
            S = S_new
        pinned += abs(S - K) < 0.10
        dist += abs(S - K)
    return pinned / n, dist / n, jumps / n

if __name__ == "__main__":
    rng = random.Random(56004)
    N = 3000
    print(
        "eta      pinned share   mean |S_close - K|   branch losses per day"
    )
    for eta in (0.0, 0.10, 0.25, 0.50, -0.10, -0.25, -0.50):
        p, d, j = run(eta, N, rng)
        print(
            "{:+5.2f}      {:7.3f}          {:7.3f}             {:6.3f}"
            .format(eta, p, d, j)
        )
    print(
        "\nReading: with dealers long the strike (eta > 0) the close clusters"
        " at the"
    )
    print(
        "strike; with dealers short it (eta < 0) the close is pushed away. A"
        " branch"
    )
    print(
        "loss is a step at which the hedging feedback has no nearby stable"
        " price, so"
    )
    print("the price must jump to a distant solution.")
# eta      pinned share   mean |S_close - K|   branch losses per day
# +0.00        0.052            1.130              0.000
# +0.10        0.075            1.065              0.000
# +0.25        0.090            1.002              0.000
# +0.50        0.132            0.922              0.000
# -0.10        0.039            1.103              0.000
# -0.25        0.018            1.217              0.000
# -0.50        0.000            1.328              0.041
#
# Reading: with dealers long the strike (eta > 0) the close clusters at the
# strike; with dealers short it (eta < 0) the close is pushed away. A branch
# loss is a step at which the hedging feedback has no nearby stable price, so
# the price must jump to a distant solution.

做市商没有头寸时(η=0\eta = 0),0.052 的收盘落在行权价 0.10 之内。做市商持有该行权价多头时,这个比例稳步上升,η=+0.50\eta = +0.50 时达 0.132——约为基线的两倍半——收盘与行权价的平均距离从 1.130 降到 0.922。做市商持有空头时,行权价变成排斥:η=−0.25\eta = -0.25 时钉住比例降到 0.018,η=−0.50\eta = -0.50 时降到 0.000。

η=−0.50\eta = -0.50 时还发生了别的事:每天有 0.041 个时间步,反馈方程失去了前一价格附近的稳定解,价格只能跳到远处的解。这是上文 κ≤−1\kappa \le -1 边界的离散时间版本——当行权价附近的伽马大到一定程度,它就在局部出现。在这个模型里,空头伽马的账簿不只是增加波动率;在集中的行权价附近,它制造跳空。

按做市商头寸 \eta 分组,模拟到期日收盘落在行权价 0.10 之内的比例(数值取自上面的程序):做市商持有多头时行权价吸引价格,持有空头时排斥价格。

证据

钉死的证据。 Ni、Pearson 与 Poteshman 记录到,在期权到期日,有挂牌期权的股票收盘价聚集在行权价上,而没有挂牌期权的股票没有这种现象;他们把它归因于期权做市商与公司自营交易员的 Delta 对冲再平衡,并部分归因于期权卖方对股价的操纵 (Ni et al. 2005)。其中对冲部分的方向与 Avellaneda-Lipkin 的多头伽马机制一致。

假说 · 尚未验证

作者的假说(尚未验证): 在指数到期日,收盘落在未平仓量最大行权价 0.1% 之内的比例,在做市商持有该行权价伽马多头的日子里高于持有空头的日子——其中做市商的头寸由带方向的成交数据测得,而不是从未平仓量推断。什么结果会证明它错:当方向被测量出来时没有差别或方向相反;若差别只在用未平仓量惯例推断方向时出现,那么起作用的是惯例而不是机制(第三章)。

小结

对冲反馈只是一行代数:令 κ=λG\kappa = \lambda G,已实现波动率被乘以 1/(1+κ)1/(1+\kappa),多头伽马使市场平静,空头伽马使其躁动,κ=−1\kappa = -1 时价格必须跳空。滞后对冲下,同一机制在短时段收益中表现为均值回复(多头伽马)或动量(空头伽马),而短时段波动率在两种情况下都上升——多头伽马的平抑作用是关于日度时段的陈述。临近到期时,同一反馈把价格钉在做市商持有多头的行权价上,把它从持有空头的行权价推开,并可能在那里制造跳空。以上全部把做市商的符号当作已知。下一章问:做市商以外的人能否知道这个符号。

参考文献

Avellaneda, Marco, and Michael D. Lipkin. 2003. “A Market-Induced Mechanism for Stock Pinning.” Quantitative Finance 3 (6): 417–25.
Baltussen, Guido, Zhi Da, Sten Lammers, and Martin Martens. 2021. “Hedging Demand and Market Intraday Momentum.” Journal of Financial Economics 142 (1): 377–403.
Frey, Rüdiger, and Alexander Stremme. 1997. “Market Volatility and Feedback Effects from Dynamic Hedging.” Mathematical Finance 7 (4): 351–74.
Ni, Sophie Xiaoyan, Neil D. Pearson, and Allen M. Poteshman. 2005. “Stock Price Clustering on Option Expiration Dates.” Journal of Financial Economics 78 (1): 49–87.
Ni, Sophie Xiaoyan, Neil D. Pearson, Allen M. Poteshman, and Joshua White. 2021. “Does Option Trading Have a Pervasive Impact on Underlying Stock Prices?” The Review of Financial Studies 34 (4): 1952–86.
Platen, Eckhard, and Martin Schweizer. 1998. “On Feedback Effects from Hedging Derivatives.” Mathematical Finance 8 (1): 67–84.

Fan, Y. (2026). Dealer Gamma Exposure, Zero-Day Options and Volatility Harvesting Dynamics: Delta-Hedging Feedback Loops, Strike Pinning, Iron Condor Tail Ruin and Variance Swap Replication [做市商伽马敞口、零日到期期权与波动率收割动力学] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090160 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).