5D Library › Implied Volatility Surfaces, Skew Premia and Quantitative Risk Control of Vertical Spreads · 中文

Chapter 6 The Ratchet, the Kelly Criterion and Risks outside the Model

“Never reaches zero” is not “safe”

The author’s framework proposes a “three-tier ratchet” for allocating profits: 30 % withdrawn, 30 % placed in a Treasury buffer, 40 % left to compound in the trading capital, and calls this “immune to ruin”. The author has confirmed that the split applies only to incremental profit above the high-water mark. This chapter separates the part of “immune to ruin” that can be proved from the part that cannot, and then shows in what sense the high-water-mark ratchet avoids the trap of a shrinking stake.

First, what can be proved. Suppose no trade can lose more than the risk capital committed to it (the maximum loss of a debit spread is known at entry), and the trader always commits a fixed fraction ff (0<f<10 < f < 1) of trading capital CnC_n as the trade’s maximum loss. Then whatever happens,

Cn+1≥(1−f)Cn⇒Cn≥(1−f)nC0>0.C_{n+1} \ge (1 – f)\,C_n \quad\Longrightarrow\quad C_n \ge (1-f)^n C_0 > 0.

Withdrawing profits can only slow growth; it does not change this lower bound. Under these two assumptions, trading capital never reaches zero in the geometric sense. The conclusion is trivial but correct.

What cannot be proved is “no risk”. First, capital can come arbitrarily close to zero: after ten straight losses only (1−f)10(1-f)^{10} is left, and once capital is smaller than the cost of one contract the trader can no longer trade. Second, the long-run growth rate depends on ff and on the withdrawal rule, and can be negative. Third, the two assumptions themselves can fail: losses can exceed the plan, and neighbouring trades can be correlated. The chapter takes these in turn.

The Kelly criterion: a brief recap

Volume 19, Non-Ergodic Wealth Dynamics, the Kelly Criterion and Volatility Drag, discussed the Kelly criterion in detail; here we only recall its conclusions. If each trade wins bb times the risk capital with probability pp and loses the risk capital with probability q=1−pq = 1 – p, betting a fraction ff gives an expected log growth per trade of

g(f)=pln⁡(1+fb)+qln⁡(1−f),g(f) = p\ln(1 + fb) + q\ln(1 – f),

maximised at f*=p−q/bf^* = p – q/b (Kelly 1956); betting f*f^* almost surely outperforms any essentially different strategy in the long run (Breiman 1961). g(f)g(f) is concave: beyond f*f^* growth declines, reaching zero near 2f*2f^* and turning negative beyond. Kelly betting also brings deep drawdowns; for Brownian motion with drift, Magdon-Ismail and co-authors derived the expected maximum drawdown analytically (Magdon-Ismail et al. 2004).

Note

The Kelly fraction requires the true pp and bb. In practice both can only be estimated from a finite record of trades, and Chapter 5 showed how large the estimation error is for a few hundred trades. Overestimating the edge pushes the bet beyond f*f^*, which is why smaller fractions such as “half Kelly” are common in practice. The values p=0.45p = 0.45 and b=1.5b = 1.5 in this chapter are illustrative.

The author’s ratchet: applied only above the high-water mark

The rule the author has confirmed, and the only reading this book adopts, is this. Call the highest value trading capital has ever reached its high-water mark HH. After a trade, only if trading capital breaks above HH is the part above HH split: 30 % withdrawn, 30 % to the Treasury buffer, 40 % left to compound. Losses are borne by trading capital, and gains that merely recover earlier losses back to the old peak stay in trading capital in full and are not skimmed.

The rule has three provable properties (propositions under the author’s definition):

  • Withdrawals come only from new highs. If a trade leaves capital at C′>HC' > H, skimming 0.6(C′−H)0.6\,(C' – H) leaves H+0.4(C′−H)>HH + 0.4\,(C' – H) > H, which becomes the new high-water mark. So after every withdrawal trading capital is still above its previous peak, and hence above the starting stake C0C_0; the total ever removed is exactly 1.5(Hn−C0)1.5\,(H_n – C_0), drawn only from new highs already reached. Withdrawals by themselves never eat into existing capital.
  • Drawdowns behave exactly as under full compounding. Because each bet is a fraction of current capital, the ratio Cn/HnC_n/H_n evolves exactly as it does with no skimming at all. The ratchet therefore does not reduce drawdowns of trading capital: the chance of losing more than half of it from a peak is the same as under full compounding.
  • The sign of the growth rate is unchanged. The high-water mark rises only at new highs, and each rise is positive. When full compounding grows, g(f)>0g(f) > 0, the drawdown ratio keeps returning to new highs, so trading capital under the high-water ratchet also has a positive long-run growth rate, though lower than g(f)g(f); when g(f)<0g(f) < 0, new highs eventually stop and capital then decays just as under full compounding. Skimming, in other words, never turns a growing strategy into a shrinking one.

That is the sense in which it avoids the trap of a shrinking stake: trading capital can shrink only through losses, never through withdrawals. It does not mean “cannot lose”: losses still cause drawdowns, trading capital can fall below the starting stake, and if the bets are too large (g(f)<0g(f) < 0) it still decays in the long run.

For contrast, and as a warning: skimming every winning trade

A reading that looks similar is to skim every winning trade: 60 % of every gain leaves trading capital, while losses are borne by trading capital in full. The book does not adopt it, but keeps it as a warning because it contains a provable trap (a proposition under the author’s definition). The log growth of trading capital per trade is then

gR(f)=pln⁡(1+0.4fb)+qln⁡(1−f).g_R(f) = p\ln(1 + 0.4fb) + q\ln(1-f).

It is concave, with gR(0)=0g_R(0) = 0 and gR′(0)=0.4pb−qg_R'(0) = 0.4pb – q. So whenever 0.4pb≤q0.4pb \le q, for every f>0f > 0 we have gR(f)≤gR′(0)f≤0g_R(f) \le g_R'(0)\,f \le 0: the long-run growth rate of trading capital is negative and it shrinks geometrically towards zero (without ever reaching it). With the illustrative parameters 0.4pb=0.27<q=0.550.4pb = 0.27 < q = 0.55, so the condition holds; for trading capital to grow under per-trade skimming the edge must be as large as pb>2.5qpb > 2.5q. The trouble is the asymmetry: losses fall on trading capital in full, while 60 % of the gains that would repair them are taken away first.

The program below compares, under the illustrative parameters, full compounding (betting f*/2f^*/2, f*f^* and 2f*2f^*), the high-water-mark ratchet adopted in this book (at the same three fractions) and, as a warning, per-trade skimming, with 20,000 paths of 250 trades each (about five years at 50 trades a year). “All wealth” includes the trading capital, the buffer (earning 3 % a year) and the amounts withdrawn; drawdowns are measured on trading capital and on “trading capital + buffer”. The program also checks the first property above on every path, and estimates the long-run growth rate of the high-water ratchet from one long path of two million trades.

Calculation

# Fixed-fraction sizing with a capped loss, the Kelly growth rate, and the author's
# three-tier ratchet: 30 % withdrawn, 30 % to a Treasury buffer at 3 %/yr, 40 % left to
# compound. ADOPTED RULE: the ratchet applies only to the part of trading capital above
# its high-water mark (HWM); a recovery back up to the old peak is never skimmed. Shown
# for contrast and as a WARNING: skimming 60 % of every winning trade. ILLUSTRATIVE
# trades: each risks a fraction f of trading capital; it loses the whole risk with
# prob. 0.55, or gains 1.5 x the risk.
import math, random

P, B, N_TRADES, PER_YEAR = 0.45, 1.5, 250, 50

def growth(f):              # expected log growth of trading capital per trade
    return P * math.log(1 + f * B) + (1 - P) * math.log(1 - f)

def run(f, ratchet, rng, n=N_TRADES):
    cap, buf, out, hwm = 1.0, 0.0, 0.0, 1.0
    peak, dd, peak2, dd2, ok = 1.0, 0.0, 1.0, 0.0, True
    for _ in range(n):
        buf *= math.exp(0.03 / PER_YEAR)
        gain = cap * f * (B if rng.random() < P else -1.0)
        skim = 0.0
        if ratchet == "trade" and gain > 0:
            skim = 0.6 * gain
        elif ratchet == "hwm" and cap + gain > hwm:
            skim = 0.6 * (cap + gain - hwm)
        out += skim / 2; buf += skim / 2; cap += gain - skim
        if ratchet == "hwm" and skim > 0:
            ok = ok and cap > hwm                  # still above the old peak after a skim
        hwm = max(hwm, cap)
        peak = max(peak, cap); dd = max(dd, 1 - cap / peak)
        peak2 = max(peak2, cap + buf); dd2 = max(dd2, 1 - (cap + buf) / peak2)
    if ratchet == "hwm":                           # all skims = 1.5 x (HWM - start)
        skimmed = out * 2                          # (buffer interest excluded)
        ok = ok and abs(skimmed - 1.5 * (hwm - 1.0)) < 1e-9 * max(1.0, hwm)
    return cap, buf, out, dd, dd2, ok

def long_run(f, ratchet, n=2000000, seed=7):
    # long-run log growth of trading capital per trade, from one long path; the state is
    # x = capital / HWM, since sizing is proportional the scale does not matter
    rng, x, lg = random.Random(seed), 1.0, 0.0
    for _ in range(n):
        y = x * (1 + f * (B if rng.random() < P else -1.0))
        if ratchet and y > 1:                  # new high: keep 40 % of the excess
            lg += math.log((1 + 0.4 * (y - 1)) / x); x = 1.0
        else:
            lg += math.log(y / x); x = min(y, 1.0) if ratchet else 1.0
    return lg / n

if __name__ == "__main__":
    k = P - (1 - P) / B
    print(f"win prob {P}, win/loss {B}: edge {P * B - (1 - P):.3f} per unit risked,"
          f" Kelly fraction f* = {k:.4f}")
    print("  f / f*     0.25   0.5    1.0    1.5    2.0    2.5")
    print("  g x 1000" + "".join(f"{1000 * growth(x * k):7.2f}" for x in
                                  (0.25, 0.5, 1.0, 1.5, 2.0, 2.5)))
    print(f"after 10 straight losses at f*: capital x {(1 - k) ** 10:.3f}"
          " (never zero, since each loss <= f* < 1)")
    print(f"{N_TRADES} trades, 20,000 paths  median: trading  all    P(drawdown >= 50%)")
    print("                                    capital  wealth   trading  trading+buffer")
    checks = []
    for lab, f, rat in (("full compounding, f*/2", k / 2, None),
                        ("full compounding, f*", k, None),
                        ("full compounding, 2 f*", 2 * k, None),
                        ("ADOPTED: above high-water, f*/2", k / 2, "hwm"),
                        ("ADOPTED: above high-water, f*", k, "hwm"),
                        ("ADOPTED: above high-water, 2f*", 2 * k, "hwm"),
                        ("WARNING: per winning trade, f*", k, "trade")):
        rng = random.Random(19); res = [run(f, rat, rng) for _ in range(20000)]
        med = lambda x: sorted(x)[len(x) // 2]
        cap = med([r[0] for r in res]); tot = med([r[0] + r[1] + r[2] for r in res])
        p50 = sum(r[3] >= 0.5 for r in res) / len(res)
        p2 = sum(r[4] >= 0.5 for r in res) / len(res)
        print(f"  {lab:31s}   {cap:7.3f}  {tot:6.2f}   {p50:6.1%}    {p2:6.1%}")
        if rat == "hwm":
            checks.append((lab.split(", ")[1], all(r[5] for r in res),
                           sum(r[0] < 1 for r in res) / len(res)))
    print("high-water ratchet, every path: after each skim capital is still above the")
    print("  old peak, and total skimmed = 1.5 x (HWM - start): "
          + ", ".join(f"{f} {'ok' if c else 'FAILED'}" for f, c, _ in checks))
    print("  but losses still bite: trading capital ends below its start in")
    print("  " + ", ".join(f"{b:.1%} of paths ({f})" for f, _, b in checks))
    g = P * math.log(1 + 0.4 * k * B) + (1 - P) * math.log(1 - k)
    print(f"per-trade ratchet: log growth of trading capital {1000 * g:+.1f} per 1000"
          f" per trade -> x {math.exp(N_TRADES * g):.4f}")
    print("long-run log growth of trading capital per 1000 trades      f*/2     f*")
    print(f"  full compounding (exact g)                              "
          f"{1000 * growth(k / 2):5.2f}  {1000 * growth(k):5.2f}")
    print(f"  ratchet above high-water (one path, 2,000,000 trades)   "
          f"{1000 * long_run(k / 2, True):5.2f}  {1000 * long_run(k, True):5.2f}")
# win prob 0.45, win/loss 1.5: edge 0.125 per unit risked, Kelly fraction f* = 0.0833
#   f / f*     0.25   0.5    1.0    1.5    2.0    2.5
#   g x 1000   2.27   3.87   5.15   3.89   0.14  -6.12
# after 10 straight losses at f*: capital x 0.419 (never zero, since each loss <= f* < 1)
# 250 trades, 20,000 paths  median: trading  all    P(drawdown >= 50%)
#                                     capital  wealth   trading  trading+buffer
#   full compounding, f*/2              2.501    2.50    36.5%     36.5%
#   full compounding, f*                3.268    3.27    97.2%     97.2%
#   full compounding, 2 f*              0.845    0.85   100.0%    100.0%
#   ADOPTED: above high-water, f*/2     1.301    2.28    36.5%     18.4%
#   ADOPTED: above high-water, f*       1.240    3.04    97.2%     70.8%
#   ADOPTED: above high-water, 2f*      0.318    2.71   100.0%     98.6%
#   WARNING: per winning trade, f*      0.001    1.41   100.0%     38.1%
# high-water ratchet, every path: after each skim capital is still above the
#   old peak, and total skimmed = 1.5 x (HWM - start): f*/2 ok, f* ok, 2f* ok
#   but losses still bite: trading capital ends below its start in
#   27.2% of paths (f*/2), 41.0% of paths (f*), 69.4% of paths (2f*)
# per-trade ratchet: log growth of trading capital -25.9 per 1000 per trade -> x 0.0015
# long-run log growth of trading capital per 1000 trades      f*/2     f*
#   full compounding (exact g)                               3.87   5.15
#   ratchet above high-water (one path, 2,000,000 trades)    1.57   2.12

The Kelly fraction is 0.0833; growth is 5.15 per thousand per trade at f*f^*, almost zero (0.14) at 2f*2f^* and negative at 2.5f*2.5f^*. After ten straight losses at f*f^*, 41.9 % of capital remains, which is not zero.

With full compounding at f*f^*, median capital after 250 trades is 3.27 times the start, but 97.2 % of paths suffer a drawdown of more than half along the way; at half Kelly the median is 2.50 and that share falls to 36.5 %; at 2f*2f^* the median is only 0.85.

The high-water ratchet adopted in this book. The program confirms that on every path, after each withdrawal, trading capital is still above its previous high-water mark, and the total withdrawn equals 1.5(Hn−C0)1.5\,(H_n – C_0). At half Kelly, median trading capital is 1.30 and total wealth 2.28 (full compounding 2.50), and the chance that “trading capital + buffer” falls by more than half drops from 36.5 % to 18.4 %. At f*f^*, median trading capital is 1.24 and total wealth 3.04 (full compounding 3.27), and that chance drops from 97.2 % to 70.8 %. But the chance of trading capital itself falling by more than half is exactly the same as under full compounding (36.5 % and 97.2 %), and after 250 trades trading capital is below its starting value on 27.2 % and 41.0 % of paths respectively, and on 69.4 % at 2f*2f^*: losses still happen. In the long run, trading capital under the high-water ratchet grows by 1.57 and 2.12 per thousand per trade at half Kelly and at f*f^*, positive but below the 3.87 and 5.15 of full compounding. The value of the ratchet lies in moving part of one’s wealth out of risk, not in raising growth.

Per-trade skimming, the warning case, confirms the proposition: trading capital grows at −25.9 per thousand per trade in log terms, and after 250 trades the median is about 0.1 % of the start (0.001 in the program, the same order as the 0.0015 given by e250gRe^{250 g_R}). Median total wealth is 1.41, almost all of it from early withdrawals and the buffer; the trading capital itself exists in name only. Figure 4 shows the Kelly growth curve and the distribution of trading capital under the three rules.

Left: expected log growth per trade against the fraction bet (in units of the Kelly fraction f^*), illustrative parameters p = 0.45, b = 1.5. Right: distribution of trading capital after 250 trades at f^* (5,000 paths each, log axis): full compounding, the high-water-mark ratchet adopted in this book, and per-trade skimming as a warning.

Risks outside the model

The calculations above assume independent trades, losses that never exceed plan and fixed costs. The program below keeps the same average win rate and bets half Kelly, then adds three departures common in real markets one by one: losses cluster in persistent “bad regimes”; the trader plans his risk with a stop at half the debit, but 10 % of losses gap through it straight to the full debit (twice the planned loss); and bid–ask spreads widen in bad regimes.

Calculation

# Risks outside the i.i.d. model, same ILLUSTRATIVE trades as p11 (win 1.5 x risk with
# prob. 0.45, else lose 1 x), sized at half Kelly. (1) losing streaks cluster in a
# persistent bad regime with the same average win rate; (2) a trader who sizes by a
# planned stop at half the debit is sometimes gapped through it to the full debit;
# (3) spreads widen in the bad regime. 250 trades, 20,000 paths each.
import math, random

P, B = 0.45, 1.5
F = 0.5 * (P - (1 - P) / B)                      # half Kelly

def path(rng, regimes=False, gap=0.0, widen=False, n=250):
    w, peak, dd, bad = 1.0, 1.0, 0.0, rng.random() < 1 / 3
    for _ in range(n):
        if regimes:                               # stay 0.95 (good) / 0.90 (bad)
            bad = (rng.random() < 0.90) if bad else (rng.random() >= 0.95)
            p = 0.25 if bad else 0.55             # long-run average still 0.45
        else:
            p = P
        if rng.random() < p:
            x = B - (0.2 if widen and bad else 0.0)
        else:
            x = -1.0 - (0.1 if widen and bad else 0.0)
            if rng.random() < gap:
                x = -2.0                          # the stop was gapped: full debit lost
        w *= 1 + F * x; peak = max(peak, w); dd = max(dd, 1 - w / peak)
    return math.log(w) / n, dd, w

if __name__ == "__main__":
    print(f"half Kelly f = {F:.4f} of capital at risk per trade")
    print("scenario                        growth x 1000  P(drawdown  wealth after 250")
    print("                                (median)       >= 50 %)    trades, 5th pct")
    for lab, kw in (("i.i.d. baseline", {}),
                    ("regimes (losses cluster)", dict(regimes=True)),
                    ("stop gapped in 10 % of losses", dict(gap=0.10)),
                    ("regimes + wider spreads", dict(regimes=True, widen=True)),
                    ("all three", dict(regimes=True, widen=True, gap=0.10))):
        rng = random.Random(20)
        res = [path(rng, **kw) for _ in range(20000)]
        g = sorted(r[0] for r in res)[10000]
        p50 = sum(r[1] >= 0.5 for r in res) / 20000
        w5 = sorted(r[2] for r in res)[1000]
        print(f"  {lab:30s}{1000 * g:8.2f}      {p50:7.1%}       {w5:6.2f}")
# half Kelly f = 0.0417 of capital at risk per trade
# scenario                        growth x 1000  P(drawdown  wealth after 250
#                                 (median)       >= 50 %)    trades, 5th pct
#   i.i.d. baseline                   3.67        36.6%         0.65
#   regimes (losses cluster)          4.08        65.8%         0.39
#   stop gapped in 10 % of losses     1.41        65.6%         0.34
#   regimes + wider spreads           2.18        79.7%         0.22
#   all three                        -0.17        89.6%         0.11

With independent trades and half Kelly, the chance of a drawdown of more than half within 250 trades is 36.6 %. When losses cluster, median growth barely changes (4.08 per thousand) but that chance rises to 65.8 %, and the fifth percentile of wealth falls from 0.65 to 0.39: the same average win rate carries far more risk. Gapped stops cut growth to 1.41 per thousand; with all three departures together, median growth turns negative (−0.17 per thousand), 89.6 % of paths lose more than half at some point, and the fifth percentile is left with 11 % of the start.

Other risks are not simulated:

  • Gaps and early assignment (Chapter 3): the maximum loss of a debit spread is capped when held to expiry, but if the short leg is assigned and the long leg is not handled in time, exposure can exceed the debit.
  • Vanishing liquidity: under market stress option bid–ask spreads can multiply or quotes can disappear altogether, so “closing on schedule” cannot be carried out.
  • Margin calls: a debit spread on its own does not face them while both legs remain open, but a trader who also holds credit spreads or written options, or trades on margin, can be forced to close at the worst moment.
  • Counterparty and platform risk: if a broker or the clearing chain fails, cash and positions in the account can be frozen.

These risks show that the two assumptions behind “never reaches zero” are exactly the ones most likely to break under stress.

Hypothesis · not established

A ratchet rule written down in advance reduces the tendency to raise bets after a losing streak (not established). The author’s hypothesis is that traders who write their withdrawal and sizing rules down in advance raise their bet fraction after a run of losses less often than traders without written rules. It is a hypothesis about behaviour, not about returns. If, in a simulated-trading experiment, groups with and without written rules show no difference in how they change their bets after losing streaks, the hypothesis is wrong.

Chapter summary

If each loss is capped and a fixed fraction is bet, trading capital never reaches zero in the geometric sense, but that does not make it safe. The Kelly fraction maximises long-run growth and brings deep drawdowns; beyond about twice the Kelly fraction growth turns negative. Following the author’s confirmation, the book applies the three-tier ratchet only to incremental profit above the high-water mark: withdrawals come only from new highs, trading capital stays above its previous peak after every withdrawal, and skimming neither eats into existing capital nor changes the sign of the growth rate; but drawdowns of trading capital are the same as under full compounding, and losses can still take it below the starting stake. Skimming every winning trade is a trap: whenever 0.4pb≤q0.4pb \le q, trading capital has a negative long-run growth rate. Clustered losses, gaps, liquidity and margin, all outside the model, raise drawdowns substantially and can make growth negative.

References

Breiman, Leo. 1961. “Optimal Gambling Systems for Favorable Games.” Proceedings of the Fourth Berkeley Symposium on Mathematical Statistics and Probability (Berkeley) 1: 65–78.
Kelly, Jr., J. L. 1956. “A New Interpretation of Information Rate.” Bell System Technical Journal 35 (4): 917–26. https://doi.org/10.1002/j.1538-7305.1956.tb03809.x.
Magdon-Ismail, Malik, Amir F. Atiya, Amrit Pratap, and Yaser S. Abu-Mostafa. 2004. “On the Maximum Drawdown of a Brownian Motion.” Journal of Applied Probability 41 (1): 147–61. https://doi.org/10.1239/jap/1077134674.

Fan, Y. (2026). Implied Volatility Surfaces, Skew Premia and Quantitative Risk Control of Vertical Spreads: Microstructure Pricing and Time-Stop Dynamics under Non-Gaussian Tails [期权隐含波动率曲面、偏度溢价与垂直价差结构的定量风控] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23796765 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).