5D Library › Macro-Liquidity Regime Switching, Extreme Tail Dependence and Barbell Asset Allocation · 中文

Chapter 8 Practice and a Pre-Registered Test

Note

The checklist and test protocol in this chapter are for research and self-examination, not investment advice. The checklist cannot tell you what to buy or sell; it can only help you write down your assumptions before you decide. The portfolios, returns and economies in the book are fictional; any real investment can lose money.

Part 1: a four-step checklist

Chapters 1 to 6 can be condensed into a four-step checklist for examining the tail risk and buffers of a portfolio (or an institution, or a country). The limits in it (for instance, “flag when ES exceeds the normal ES by more than 20 %”) are examples of rules the user should write down in advance, not recommended values.

  1. Regime. Using information available in real time (filtered probabilities, not smoothed ones), which regime does the present look like? How many regime switches does the estimate of the regime parameters rest on? With too few, estimates of durations and means mean almost nothing (Chapter 1).
  2. Tail. How fat is the tail on the Hill plot and the mean-excess plot? How much higher is the ES estimated by extreme value methods than the ES under normality? Does the conclusion survive a change of threshold (Chapter 2)?
  3. Dependence. On the worst 5 % of days, what is the conditional probability that the assets fall together? How much higher is it than for a Gaussian copula with the same pairwise rank correlations (Chapters 3 and 4)?
  4. Buffer. Can safe assets available immediately cover normal needs plus stressed needs re-estimated from the tail? If the buffer ratio is below 1, where will the shortfall come from (Chapters 5 and 6)?

The program below applies the checklist to a fictional two-asset portfolio (2,500 simulated trading days, the last 40 of them in a stressed regime).

Calculation

# The four-step checklist applied to a FICTIONAL two-asset portfolio (2,500 simulated
# days, returns in %): (1) regime: filtered P(stress) today from a two-state filter on
# asset 1 whose parameters are set by a written rule, not fitted; (2) tails: Hill index
# and GPD 99 % ES against the normal ES; (3) dependence: lower-tail co-exceedance at the
# 5 % level against a Gaussian copula with the same rank correlation; (4) buffer: safe
# liquid assets against a stressed 12-month need. Flags use thresholds written down
# BEFORE looking.
import math, random, statistics

ND = statistics.NormalDist()

def make_data(rng, T=2500):
    s, x, y, z = 0, [], [], []
    for t in range(T):
        sd = 0.8 if s == 0 else 2.0
        crash = s == 1 and rng.random() < 0.04
        f = rng.gauss(0, 1) - (3.0 if crash else 0.0)
        x.append(sd * (0.7 * f + 0.71 * rng.gauss(0, 1)) / 1.0)
        y.append(sd * (0.7 * f + 0.71 * rng.gauss(0, 1)) / 1.0)
        z.append(s)
        stay = 0.995 if s == 0 else 0.97
        if t < T - 40:
            s = s if rng.random() < stay else 1 - s
        else:
            s = 1                                     # the last 40 days are stressed
    return x, y, z

def regime(x, p_stay=(0.99, 0.97)):     # two-state forward filter; parameters set by a
    s0 = 1.4826 * statistics.median(abs(v) for v in x)  # rule fixed in advance: calm sd
    sd = (s0, 2.5 * s0); a, spells = [1.0, 0.0], 0      # from the median |return|
    for v in x:
        pr = [a[0] * p_stay[0] + a[1] * (1 - p_stay[1]),
              a[0] * (1 - p_stay[0]) + a[1] * p_stay[1]]
        f = [pr[k] * math.exp(-0.5 * (v / sd[k]) ** 2) / sd[k] for k in range(2)]
        new = [f[0] / sum(f), f[1] / sum(f)]; spells += new[1] > 0.5 >= a[1]; a = new
    return a[1], sd, spells

def hill(loss, k):
    s = sorted(loss, reverse=True)
    return k / sum(math.log(s[i] / s[k]) for i in range(k))

def gpd_es(loss, k=125, p=0.99):                       # probability-weighted moments
    s = sorted(loss, reverse=True); u = s[k]; y = sorted(v - u for v in s[:k])
    b0 = sum(y) / k; b1 = sum((k - 1 - i) / (k - 1) * v for i, v in enumerate(y)) / k
    xi = 2 - b0 / (b0 - 2 * b1); sig = 2 * b0 * b1 / (b0 - 2 * b1)
    var = u + sig / xi * ((len(loss) / k * (1 - p)) ** -xi - 1)
    return (var + sig - xi * u) / (1 - xi)

def co_exceed(x, y, u=0.05):
    k = int(u * len(x)); qx, qy = sorted(x)[k], sorted(y)[k]
    return sum(1 for a, b in zip(x, y) if a < qx and b < qy) / k

def spearman(x, y):
    rx = {v: i for i, v in enumerate(sorted(x))}
    ry = {v: i for i, v in enumerate(sorted(y))}
    return statistics.correlation([rx[v] for v in x], [ry[v] for v in y])

if __name__ == "__main__":
    rng = random.Random(40); x, y, z = make_data(rng)
    port = [(a + b) / 2 for a, b in zip(x, y)]; loss = [-v for v in port]
    p, sd2, spells = regime(x)
    print(f"1 regime    filtered P(stress | data to today) = {p:.2f}"
          f" (calm sd {sd2[0]:.2f}, stress sd {sd2[1]:.2f})")
    print(f"            flag: P > 0.5 -> {'YES' if p > 0.5 else 'no'}")
    m, sd = statistics.fmean(loss), statistics.stdev(loss)
    es_n = m + sd * math.exp(-ND.inv_cdf(0.99) ** 2 / 2) / math.sqrt(2 * math.pi) / 0.01
    es_g = gpd_es(loss)
    print(f"2 tails     Hill alpha (k = 100 / 200): {hill(loss, 100):.1f} /"
          f" {hill(loss, 200):.1f}; 99 % ES: GPD {es_g:.2f}, normal {es_n:.2f}")
    flag = "YES" if es_g > 1.2 * es_n else "no"
    print(f"            flag: GPD ES > 1.2 x normal ES -> {flag}")
    rs = spearman(x, y); rho = 2 * math.sin(math.pi * rs / 6)
    sim = []
    for _ in range(20_000):
        a = rng.gauss(0, 1)
        sim.append((a, rho * a + math.sqrt(1 - rho * rho) * rng.gauss(0, 1)))
    ce, cg = co_exceed(x, y), co_exceed([s[0] for s in sim], [s[1] for s in sim])
    print(f"3 dependence  Spearman {rs:.2f}; P(both in worst 5 % | one is):"
          f" data {ce:.2f}, Gaussian {cg:.2f}")
    flag = "YES" if ce > cg + 0.10 else "no"
    print(f"            flag: data > Gaussian + 0.10 -> {flag}")
    safe, need, add_n = 18.0, 12.0, 6.0          # fictional, in months of obligations
    add = add_n * es_g / es_n                     # stress add-on scaled by the ES ratio
    ratio = safe / (need + add)
    print(f"4 buffer    safe liquid assets {safe:.0f}; need {need:.0f} + stress add-on"
          f" {add_n:.0f} x {es_g / es_n:.2f} = {need + add:.1f} -> ratio {ratio:.2f}")
    print(f"            flag: ratio < 1 -> {'YES' if ratio < 1 else 'no'}")
    print(f"(truth, hidden from the checklist: today is a stress day: {z[-1] == 1})")
# 1 regime    filtered P(stress | data to today) = 0.23 (calm sd 0.91, stress sd 2.28)
#             flag: P > 0.5 -> no
# 2 tails     Hill alpha (k = 100 / 200): 2.7 / 2.1; 99 % ES: GPD 4.40, normal 2.78
#             flag: GPD ES > 1.2 x normal ES -> YES
# 3 dependence  Spearman 0.48; P(both in worst 5 % | one is): data 0.47, Gaussian 0.23
#             flag: data > Gaussian + 0.10 -> YES
# 4 buffer    safe liquid assets 18; need 12 + stress add-on 6 x 1.58 = 21.5 -> ratio 0.84
#             flag: ratio < 1 -> YES
# (truth, hidden from the checklist: today is a stress day: True)

Step 1 gives no signal: the two-state filter, whose parameters are set by a rule written in advance (calm volatility derived from the median absolute return, stressed volatility 2.5 times that), puts the current probability of stress at only 0.23, although in truth the last 40 days have all been stressed. The filter’s parameters differ from the true ones (2.28 against 2.0 for the stressed regime) and the stressed period contains relatively quiet days, so the real-time probability is less reliable than in Chapter 1’s ideal case with true parameters; this is the practical consequence of “regimes are clear only in hindsight”. The other three steps all give signals: the Hill index of the portfolio’s losses is about 2.1 to 2.7, and the 99 % ES estimated by extreme value methods is 4.40, 58 % above the normal 2.78; on the worst 5 % of days, the conditional probability of both assets falling together is 0.47, against 0.23 for a Gaussian copula with the same rank correlation; and after the stressed need is scaled up by the tail ratio, the buffer ratio is 0.84, below 1. The lesson of the example is that the checklist’s value lies in its steps confirming one another; it should not rest on the regime signal alone.

Part 2: a pre-registered backtest

Question and design choice

On a set of daily return series fixed in advance and a locked out-of-sample window, is the violation rate of a one-day 99 % VaR from conditional EVT (GARCH plus a generalized Pareto distribution) lower than that from the conditional normal method (the same GARCH plus a normal quantile), by at least the smallest effect of interest?

The author’s framework offered two options: an out-of-sample test of “rebalancing driven by HMM regime signals versus fixed proportions”, or a backtest comparing “EVT-based ES risk control with normal VaR”. The book chooses the latter, for three reasons. First, the outcomes of the former (differences in Sharpe ratio or drawdown) are extremely noisy: Chapter 8 of Book 38 explained that maximum drawdown, a single statistic of a whole path, has almost no statistical power and that the Sharpe ratio depends on the position-sizing rule, so both served only as secondary outcomes there. The latter works at the level of “did this series violate on this day”, which provides far more data. Second, the latter tests the central claim of Chapter 2 directly, with a clear theoretical expectation and published results for comparison (McNeil and Frey 2000). Third, it involves no trading and puts no one’s money at risk. Both models use the same GARCH filter, so the comparison isolates the modelling of the tail.

Protocol

Element Specification
Data and window A third-party researcher unconnected with the author obtains daily closing data from a commercial database; the out-of-sample window is written down and locked when the protocol is registered (about 9.2 years, 2,329 trading days, as required by the sample size), with a further 1,000 trading days before the window for initial estimation; the third party releases no results for the window before the protocol is registered and the analysis code is public
Series and inclusion criteria 20 series (several each of equity indices, government bond indices, major exchange rates and commodity indices), listed at registration by data-completeness and liquidity criteria applied before the window; series-days with more than 5 % missing data are excluded
The two models A: GARCH(1,1) by normal quasi-maximum likelihood on a rolling 1,000-day window, a GPD fitted to the 100 largest standardised losses, giving the 99 % quantile (the McNeil–Frey method); B: the same GARCH with the normal quantile 2.326. Both re-estimated every 250 trading days, with GARCH volatility updated daily
Blinding No random allocation is needed (both models are computed for every series-day); the analysis code labels the two columns “model X / model Y”, and the third party reveals which is which only after the results are locked
Primary outcome DD = pooled violation rate of model B minus that of model A (percentage points), pooled over the 20 series; 95 % confidence interval from a bootstrap that resamples whole days (all 20 series of a day together)
Pre-specified qualifier Whether model A’s own pooled violation rate lies within 0.7–1.3 %; outside that range, the verdict carries the qualifier “over-conservative” or “also under-covers”
Secondary outcomes Kupiec’s test and Christoffersen’s independence and conditional-coverage tests for each model, per series and pooled (Kupiec 1995; Christoffersen 1998); the ratio of actual loss to forecast ES on violation days; the skill score in quantile (tick) loss; a joint score for 97.5 % ES (Fissler and Ziegel 2016); unconditional normal and historical simulation as further comparisons; all descriptive or exploratory

There is a single primary outcome, the “difference in violation rates”, for these reasons: it corresponds directly to the claim of Chapter 2 (normal methods understate the tail and are violated too often); the two models are paired on the same series-days, removing differences in market conditions; and the result is easy to interpret. In theory a better comparison is quantile loss (a scoring function “consistent” for quantiles (Gneiting 2011)), but the program shows that it is very insensitive to tail errors and has too little power (see the sample size below), so it is only a secondary outcome. The weakness of a difference in violation rates is that an over-conservative model also has a low violation rate; the protocol therefore pre-specifies a qualifier checking whether model A’s own rate is close to the nominal 1 %.

Guarding against data snooping

As in Book 38: there is a single, pre-fixed model specification and a single primary outcome, and no parameters are tuned (the threshold is fixed at 100 exceedances and the window at 1,000 days); the data, window, list of series and analysis code are registered publicly before the third party obtains the data, with a hash of the code registered alongside the protocol (Nosek et al. 2018); the third party runs all the code; secondary and exploratory results are clearly labelled and cannot replace the primary conclusion. Book 33 discussed the statistical principles behind these practices.

Sample size

Calculation

# Sample size for the Chapter 8 pre-registered backtest: GARCH-EVT against GARCH-normal
# one-day 99 % VaR. Both models share the same GARCH volatility, so they differ only in
# the quantile of the standardized loss z. Primary outcome: D = violation rate of the
# normal model minus that of the EVT model (percentage points), paired on the same
# series-days.
# ASSUMED, not estimated from data: z is Student t with 5 d.f. (unit variance); the EVT
# quantile is re-estimated every 250 days from 1,000 residuals (GPD on the top 100);
# 20 series; intraclass correlation (ICC) of the paired difference across the 20 series on
# the same day 0.2; 252 trading days a year; 10 % of series-days lost.
import math, random, statistics

ND = statistics.NormalDist(); Z = ND.inv_cdf
DF = 5; SC = math.sqrt(DF / (DF - 2))

def tdraw(rng):
    return rng.gauss(0, 1) / math.sqrt(rng.gammavariate(DF / 2, 2) / DF) / SC

def gpd_q(z, k=100, p=0.99):                   # probability-weighted-moment GPD fit
    s = sorted(z, reverse=True); u = s[k]; y = sorted(v - u for v in s[:k])
    b0 = sum(y) / k; b1 = sum((k - 1 - i) / (k - 1) * v for i, v in enumerate(y)) / k
    xi = 2 - b0 / (b0 - 2 * b1); sig = 2 * b0 * b1 / (b0 - 2 * b1)
    return u + sig / xi * ((len(z) / k * (1 - p)) ** -xi - 1)

def tick(z, q, tau=0.99):
    return (z - q) * (tau - (1.0 if z < q else 0.0))

def years(sesoi, sd, icc=0.2, m=20, power=0.90, lost=0.10):
    n_eff = ((Z(0.975) + Z(power)) * sd / sesoi) ** 2
    deff = 1 + (m - 1) * icc
    return n_eff, deff, n_eff * deff / m / 252 / (1 - lost)

if __name__ == "__main__":
    rng = random.Random(1995); qn = Z(0.99); vn = ve = up = dn = 0; tn, te = [], []
    for _ in range(400):                       # 400 blocks of 250 days = 100,000 days
        qe = gpd_q([tdraw(rng) for _ in range(1000)])
        for _ in range(250):
            z = tdraw(rng); a, b = z > qn, z > qe
            vn += a; ve += b; up += a and not b; dn += b and not a
            tn.append(tick(z, qn)); te.append(tick(z, qe))
    n = 100_000; d = (vn - ve) / n; sd = math.sqrt((up + dn) / n - d * d)
    print(f"violation rates: normal {vn / n:.2%}, EVT {ve / n:.2%};"
          f" expected D = {100 * d:.2f} points")
    print(f"   (SD of the paired difference per series-day {sd:.3f})")
    n_eff, deff, yrs = years(0.0025, sd)
    print(f"SESOI 0.25 points, 90 % power, alpha 0.05 (two-sided): {n_eff:,.0f}"
          f" independent series-days")
    print(f"   design effect {deff:.1f} -> {n_eff * deff / 20:,.0f} days per series"
          f" (20 series) = {n_eff * deff / 20 / 252:.1f} years;")
    print(f"   with 10 % lost: {yrs:.1f} years ({math.ceil(yrs * 252):,} trading days)")
    print("years of data (20 series, 10 % lost)    SESOI 0.15   0.25   0.40 points")
    for icc in (0.1, 0.2, 0.4):
        print(f"   ICC {icc:.1f}{'':31s}" + "".join(f"{years(s, sd, icc)[2]:7.1f}"
                                                  for s in (0.0015, 0.0025, 0.0040)))
    se = 0.0025 / (Z(0.975) + Z(0.90))
    print(f"if the true D equals the SESOI: P(lower bound > 0) ="
          f" {1 - ND.cdf((1.96 * se - 0.0025) / se):.3f};")
    print(f"   P(lower bound >= SESOI) = {1 - ND.cdf(1.96):.3f}")
    ss = 1 - statistics.fmean(te) / statistics.fmean(tn)
    dd = statistics.stdev(a - b for a, b in zip(tn, te)) / statistics.fmean(tn)
    print(f"why not the tick (quantile) loss? expected skill score of EVT {ss:+.3f};")
    print(f"   detecting even 0.03 would take {years(0.03, dd)[2]:.1f} years"
          f" of 20 series")
# violation rates: normal 1.54%, EVT 1.03%; expected D = 0.51 points
#    (SD of the paired difference per series-day 0.072)
# SESOI 0.25 points, 90 % power, alpha 0.05 (two-sided): 8,732 independent series-days
#    design effect 4.8 -> 2,096 days per series (20 series) = 8.3 years;
#    with 10 % lost: 9.2 years (2,329 trading days)
# years of data (20 series, 10 % lost)    SESOI 0.15   0.25   0.40 points
#    ICC 0.1                                  15.5    5.6    2.2
#    ICC 0.2                                  25.7    9.2    3.6
#    ICC 0.4                                  46.0   16.6    6.5
# if the true D equals the SESOI: P(lower bound > 0) = 0.900;
#    P(lower bound >= SESOI) = 0.025
# why not the tick (quantile) loss? expected skill score of EVT +0.010;
#    detecting even 0.03 would take 11.4 years of 20 series

The assumptions (not estimates from real data) are that standardised losses are Student t with 5 degrees of freedom; under that assumption the normal quantile is violated at a rate of 1.54 % and the EVT quantile at 1.03 %, an expected difference DD of 0.51 percentage points. For comparison, McNeil and Frey’s results for the S&P 500 (104 against 73 violations in about 7,400 days) correspond to about 0.42 points. The smallest effect of interest is 0.25 percentage points, a quarter of the nominal violation rate: a smaller difference would not justify changing a risk model. With two-sided α=0.05\alpha = 0.05 and 90 % power, 8,732 independent series-days are needed; the intraclass correlation of the paired differences across the 20 series on the same day is assumed to be 0.2, a design effect of 4.8, so each series needs 2,096 trading days (about 8.3 years); allowing for 10 % data loss, the plan is 2,329 trading days, about 9.2 years. With the 1,000-day initial estimation window, each series needs about 13 years of data in all.

The sensitivity table shows that the design depends most on the intraclass correlation: crashes often hit many markets at once, so violations on the same day are highly correlated. With an intraclass correlation of 0.4, 16.6 years are needed; with a smallest effect of interest of 0.15 points and an intraclass correlation of 0.2, 25.7 years. The last line of the program explains why quantile loss is not the primary outcome: under the same assumptions the expected skill score of EVT is only +0.010 (a 1 % reduction in quantile loss); even detecting an improvement of 0.03 would take 11.4 years, and detecting 0.010 about nine times as much data.

Following Chapter 6 of Book 33, “90 % power” means that if the true effect equals the smallest effect of interest exactly, the probability of reaching some conclusion that “EVT is violated less often” (a confidence interval whose lower bound is above 0) is about 90 %; in that case the strongest conclusion, “at least 0.25 points fewer”, has only about a 2.5 % chance of appearing. Reaching the strongest conclusion with high probability requires a true effect clearly larger than 0.25 points.

Decision rule

Calculation

# Decision rule, fixed before the locked data window is opened. D = pooled 99 % VaR
# violation rate of GARCH-normal minus that of GARCH-EVT (percentage points), with a 95 %
# CI from a bootstrap that resamples whole trading days (all 20 series together).
# SESOI 0.25 points. Gates (checked first): more than 15 % of series-days lost; the
# independent re-run of the forecasting code disagreeing on more than 1 % of forecasts;
# more than 5 % of GARCH refits failing to converge -> NOT INTERPRETABLE. A pre-specified
# qualifier reports whether the EVT model's own pooled rate lies within 0.7-1.3 %.
SESOI = 0.25

def verdict(d, lo, hi, evt_rate, lost=0.04, mismatch=0.001, nonconv=0.01):
    for x, lim, why in ((lost, 0.15, "series-days lost"),
                        (mismatch, 0.01, "forecasts not reproduced"),
                        (nonconv, 0.05, "GARCH refits failed")):
        if x > lim:
            return f"D = {d:+.2f}: NOT INTERPRETABLE - {x:.0%} {why} (limit {lim:.0%})"
    if lo >= SESOI:   v = "normal VaR violated at least 0.25 points more often"
    elif lo > 0:      v = "normal VaR violated more often; size not shown to reach 0.25"
    elif hi < 0:      v = "EVT VaR violated more often than normal VaR"
    elif lo >= -SESOI and hi <= SESOI: v = "no difference worth acting on"
    else:             v = "unclear: the interval is too wide to decide"
    if evt_rate < 0.7:
        v += f"\n    qualifier: EVT is over-conservative ({evt_rate:.2f} %)"
    elif evt_rate > 1.3:
        v += f"\n    qualifier: EVT also under-covers ({evt_rate:.2f} %)"
    return f"D = {d:+.2f} ({lo:+.2f} to {hi:+.2f}): {v}"

for case in ((0.52, 0.31, 0.73, 1.04), (0.20, 0.06, 0.34, 0.98),
             (0.04, -0.12, 0.20, 1.10), (-0.30, -0.49, -0.11, 1.42),
             (0.15, -0.20, 0.50, 1.00), (0.61, 0.38, 0.84, 0.55),
             (0.40, 0.20, 0.60, 1.02, 0.22)):
    print(verdict(*case))
print("secondary outcomes (Kupiec and Christoffersen tests for each model, the tick-loss")
print("  skill score, the ES backtest, the unconditional-normal model) are reported with")
print("  intervals; they never decide the verdict")
# D = +0.52 (+0.31 to +0.73): normal VaR violated at least 0.25 points more often
# D = +0.20 (+0.06 to +0.34): normal VaR violated more often; size not shown to reach 0.25
# D = +0.04 (-0.12 to +0.20): no difference worth acting on
# D = -0.30 (-0.49 to -0.11): EVT VaR violated more often than normal VaR
#     qualifier: EVT also under-covers (1.42 %)
# D = +0.15 (-0.20 to +0.50): unclear: the interval is too wide to decide
# D = +0.61 (+0.38 to +0.84): normal VaR violated at least 0.25 points more often
#     qualifier: EVT is over-conservative (0.55 %)
# D = +0.40: NOT INTERPRETABLE - 22% series-days lost (limit 15%)
# secondary outcomes (Kupiec and Christoffersen tests for each model, the tick-loss
#   skill score, the ES backtest, the unconditional-normal model) are reported with
#   intervals; they never decide the verdict

The example results are made up. If any precondition fails (more than 15 % of series-days lost, the third party’s independent rerun of the forecasts disagreeing with the registered code’s results on more than 1 % of forecasts, or more than 5 % of GARCH re-estimations failing to converge), the report says “not interpretable” whatever the primary outcome. Only when the lower bound of the confidence interval is at least 0.25 does the report say “normal VaR is violated at least 0.25 points more often”; when the lower bound is above 0 but below 0.25, “normal VaR is violated more often, but the size is not established”; when the upper bound is below 0, “EVT is violated more often”; when the interval lies within ±0.25, “no difference worth changing the model for”; otherwise, “cannot decide”. The qualifier does not change the category of the verdict but is appended to it: in the sixth example, for instance, the EVT model’s violation rate is only 0.55 %, and the report must also say that it is “over-conservative”.

What this test cannot show

First, historical results do not guarantee future ones; the tail behaviour of markets changes with regimes, institutions and market structure. Second, it tests one specific setting (GARCH(1,1), a 1,000-day window, 100 exceedances, a one-day 99 % VaR) and cannot be generalised to all extreme value methods or all levels. Third, a violation rate says how often the line is crossed, not how large the losses beyond it are; their size is described by the ES comparison among the secondary outcomes. Fourth, fewer violations may simply mean more conservatism, so the qualifier and Kupiec’s test must be read alongside. Fifth, 20 series and one time window cannot represent all markets and periods, and correlation between series makes the effective sample much smaller than the nominal one. Sixth, it does not test the copula and vine models of Chapters 3 and 4, the barbell of Chapter 5 or the regime signals of Chapter 1; those need their own tests. Seventh, the author may have lived through the market history in the window, so it is not strictly “unseen” data; a prospective replication on data arriving after the protocol is registered would be stronger evidence.

Ethics and conflicts of interest

The test involves no human participants and needs no ethics committee approval; but the protocol, list of series, model specifications and analysis code must be registered publicly before the third party obtains the data. Data are used under the vendor’s licence terms. Whatever the result, positive or negative, the results, the code and (where the licence allows) summary data will be published.

Disclosure

The author uses the works of 5D Library, including this book, as the methodological basis of his community courses, the 5D Ascension Cards (Dimensional Shift Card Decks) and the Yafu AI knowledge base, and therefore has a commercial interest in the subject of this book. The book promotes no course, card deck or service; the Chapter 8 trial is designed to be run and analysed by researchers independent of the author’s business and does not recruit from the author’s community, students or customers; all the code and the study protocol are published openly for anyone to check. The author develops and maintains quantitative research scripts and provides related open-source research and independent consulting services; while writing this book he may have held routine observation positions in US equity index funds (such as SPY) or related derivatives. All models in this book are offered solely for mathematical and historical-statistical discussion; they constitute no investment advice or trading signal, and the book recommends no specific trade or asset. To limit the influence of these interests, the protocol and analysis code are registered publicly before the third party obtains the data, and the results and code will be published whatever the outcome.

Chapter summary

The four-step checklist examines tail risk from four sides, regime, tail, dependence and buffer; the example shows that the regime signal can fail in real time and that the other steps need to confirm one another. The pre-registered backtest compares the one-day 99 % VaR violation rates of conditional EVT and the conditional normal method on 20 series and about 9.2 years of locked out-of-sample data, with a single primary outcome, the difference between them, a smallest effect of interest of 0.25 percentage points, execution by a third party, a decision rule written as code and a pre-specified “over-conservative” qualifier; it will be published whatever the result. The test examines one specific extreme value setting and cannot prove the book’s other models.

References

Christoffersen, Peter F. 1998. “Evaluating Interval Forecasts.” International Economic Review 39 (4): 841–62. https://doi.org/10.2307/2527341.
Fissler, Tobias, and Johanna F. Ziegel. 2016. “Higher Order Elicitability and Osband’s Principle.” The Annals of Statistics 44 (4): 1680–707. https://doi.org/10.1214/16-AOS1439.
Gneiting, Tilmann. 2011. “Making and Evaluating Point Forecasts.” Journal of the American Statistical Association 106 (494): 746–62. https://doi.org/10.1198/jasa.2011.r10138.
Kupiec, Paul H. 1995. “Techniques for Verifying the Accuracy of Risk Measurement Models.” The Journal of Derivatives 3 (2): 73–84. https://doi.org/10.3905/jod.1995.407942.
McNeil, Alexander J., and Rüdiger Frey. 2000. “Estimation of Tail-Related Risk Measures for Heteroscedastic Financial Time Series: An Extreme Value Approach.” Journal of Empirical Finance 7 (3–4): 271–300. https://doi.org/10.1016/S0927-5398(00)00012-8.
Nosek, Brian A., Charles R. Ebersole, Alexander C. DeHaven, and David T. Mellor. 2018. “The Preregistration Revolution.” Proceedings of the National Academy of Sciences 115 (11): 2600–2606. https://doi.org/10.1073/pnas.1708274114.

Fan, Y. (2026). Macro-Liquidity Regime Switching, Extreme Tail Dependence and Barbell Asset Allocation: Vine Copulas, Hidden Markov Models and Sovereign Buffer Dynamics [宏观流动性体制转换、极值尾部相依与杠铃资产配置] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23828049 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).