TradeScript - Quant retweeted
BREAKING: The SpaceX, $SPCX, IPO will be quoted at 9:50 AM ET today and begin trading at 10:00 AM ET.
Currently, the stock is indicated to open ~25% higher, making SpaceX the 7th largest public company in the world and Elon Musk the first trillionaire in history.
TradeScript - Quant retweeted
DeSantis: “If I go to Best Buy and buy a TV, I pay a sales tax but I don’t keep paying taxes on it year after year.”
Florida Governor Ron DeSantis On Sales Tax
.@RonDeSantis : “If I go to Best Buy and buy a TV, I pay a sales tax but I don’t keep paying taxes on it year after year.”
TradeScript - Quant retweeted
A year ago: we got Buy the dip signal
S&P 500 went 40% higher
Nasdaq went 50% higher
You better not miss out this time.
When the President says buy: You buy
TradeScript - Quant retweeted
🚨 “If Iran doesn’t FULLY OPEN, WITHOUT THREAT, the Strait of Hormuz, within 48 HOURS from this exact point in time, the United States of America will hit and obliterate their various POWER PLANTS, STARTING WITH THE BIGGEST ONE FIRST…” - President DONALD J. TRUMP
Trump demands Japan deploy Godzilla to the Strait of Hormuz. Japan politely explains Godzilla is fictional. Trump pivots to King Kong via Skull Island.
Peak geopolitics.
newsthump.com/2026/03/20/tru…
51
Quant Mechanics #20: Tail Risk, CVaR, Copulas, and the Trades You Didn't Model
VIX hit 25.09 today, up 12.16% on the session, while SPY dropped 1.40% to 661.43. That's not a crash. It's a reminder that the distribution you modeled last month is not the one you were living in today.
Every risk model you run has a blind spot baked into its foundation. Value at Risk, the number your risk manager quotes on Monday mornings, tells you the minimum loss you'll suffer on your worst 1% of days. It says nothing about how bad that worst 1% actually gets. When markets really break, that silence costs you everything.
Why VaR Fails at the Worst Moment
VaR is exactly the wrong tool for the problem it's supposed to solve. The 99% 1-day VaR for SPY, given current VIX at 25.09 and annualized vol of roughly 25%, works out to:
Daily vol = 25% / sqrt(252) = 1.575%
99% VaR = 1.575% * 2.326 = 3.66% = $24.22 per share
That $24.22 number means: on your worst 1 out of 100 trading days, you should expect to lose at least this much. It says absolutely nothing about how much you'll actually lose on those days. The distribution could have you averaging $25 past that threshold, or $80.
There's also the subadditivity problem. A coherent risk measure must satisfy Risk(A + B) <= Risk(A) + Risk(B). Combining two portfolios should never increase total risk if they aren't perfectly correlated. VaR violates this. You can construct two positions where VaR(A + B) > VaR(A) + VaR(B), meaning your aggregated risk metric says combining them made things worse. This is not a theoretical curiosity. It was a practical failure during 2008, when correlation broke and concentrated bets across different asset classes blew up simultaneously.
The Basel Committee moved to Expected Shortfall in the Basel III/IV framework for exactly this reason. ES is subadditive. VaR is not.
CVaR and Expected Shortfall: The Math Behind the Tail
Expected Shortfall (ES), also called Conditional Value at Risk (CVaR), asks: given that we're in the worst alpha% of outcomes, what is the expected loss?
Formally, for a return distribution R:
ES_alpha = -E[R | R <= VaR_alpha]
For a normal distribution with daily mean mu and daily standard deviation sigma, this has a closed form:
ES_alpha = -mu + sigma * phi(z_alpha) / (1 - alpha)
Where phi is the standard normal PDF and z_alpha = Phi_inverse(alpha). At the 99% level:
z_0.99 = 2.326
phi(2.326) = 0.02665
ES_99 = -0.04% + 1.575% * 0.02665 / 0.01 = 4.16% per day = $27.51 per share
Compare that to VaR of $24.22. The gap looks modest. It won't look modest when you swap out the normal assumption for what SPY returns actually behave like.
Computing ES from historical data requires no distributional assumption at all:
[Python: historical Expected Shortfall via simulation — 99% confidence, SPY at 661.43 with VIX-implied daily vol of 1.575%]
Under normality the ES/VaR ratio sits around 1.13. In actual SPY data that ratio runs 1.4 to 1.6, because the left tail of equity returns is fat. The normal assumption underprices tail risk by 25 to 40%, consistently, year after year.
Fat Tails: The Excess Kurtosis Problem
Normal distribution has kurtosis of 3. SPY daily returns over the past decade have excess kurtosis somewhere between 4 and 8, depending on the window. The tails are 2x to 3x fatter than the normal model predicts.
The Student-t distribution handles this properly. Returns following a Student-t with nu degrees of freedom have excess kurtosis of 6/(nu - 4) for nu > 4. With nu = 5, excess kurtosis = 6/(5-4) = 6. With nu = 6, it drops to 3. Below nu = 4, kurtosis is infinite, which is actually a reasonable description of some equity crash periods.
[Python: Student-t MLE fitting and Expected Shortfall calculation using standardised left-tail quantile — ES/VaR ratio comparison at 99% confidence]
With nu around 4 to 6, typical for SPY, the ES/VaR ratio jumps to 1.5 to 2.0. At VIX = 25.09 with daily vol of 1.575%, a Student-t(5) model gives a 99% ES of roughly 5.8%, versus the normal model's 4.16%. That's a $38 expected loss on SPY in the conditional tail, not $27.
If you're sizing to a normal-distribution VaR, you're carrying 40 to 70% more tail exposure than your model says.
Copulas: The Correlation That Breaks When You Need It
Single-asset tail risk is manageable with better distributions. Multi-asset tail risk requires copulas, and this is where most quants are still flying blind.
Sklar's theorem states that any joint distribution H(x, y) can be written as:
H(x, y) = C(F(x), G(y))
Where F and G are the marginal CDFs and C is a copula on [0,1]^2. The copula captures the dependence structure entirely separately from the marginals. You can have Student-t marginals with a Gaussian copula, or normal marginals with a Clayton copula. The copula is the model for how the assets move together, independent of what each individual asset's distribution looks like.
The Gaussian copula assumes that after transforming each asset's returns to uniform marginals via their CDF, the joint structure is multivariate normal. Computationally convenient. Widely used in structured credit from 2000 to 2008. The problem is that the Gaussian copula has zero tail dependence. As both assets approach extreme losses simultaneously, the probability of joint crashes goes to zero asymptotically. The correlation that the Gaussian copula captures is a middle-of-the-distribution phenomenon. It extrapolates that correlation to the tails where it no longer holds.
The Clayton copula is asymmetric. Lower tail dependence, upper tail independence. Two assets modeled with Clayton are more likely to crash together than to spike together. This matches equity returns: when markets sell off, cross-asset correlations converge toward 1. When they rally, correlations spread out. Clayton captures that asymmetry. The Gaussian copula ignores it entirely.
[Python: Gaussian vs Clayton copula bivariate simulation — 100K paths, SPY/QQQ with rho=0.92 and theta=3.0, joint 2% tail event frequency comparison]
Typical output: Gaussian copula gives joint 2% tail events at roughly 0.016, about 40x the independence baseline of 0.0004. Clayton copula with theta = 3 gives around 0.028 to 0.032. That's a factor of 2 difference in your modeled joint tail exposure. When SPY drops 2%, QQQ drops 2% nearly twice as often as your Gaussian copula predicts.
Lower Tail Dependence Coefficient
The Clayton copula's lower tail dependence coefficient is:
lambda_L = 2^(-1/theta)
At theta = 3: lambda_L = 2^(-1/3) = 0.794
As both assets approach their extreme left tails, the conditional probability that one is in its tail given the other is also in its tail converges to 0.794. Under the Gaussian copula with rho = 0.92, lambda_L = 0. The Gaussian says joint crashes become asymptotically independent. Clayton says they become nearly certain.
The 2008 CDO market collapse was a direct consequence of this. Every synthetic CDO tranche was priced assuming default correlations lived in the middle of the distribution and dissipated at the extremes. When housing prices fell simultaneously across all geographies, the Gaussian copula underestimated joint defaults by an order of magnitude. Entire tranches rated AAA were structurally mispriced because the dependence model was wrong.
You don't need to be in structured credit for this to matter. Any long equity portfolio using diversification across sectors is making an implicit Gaussian copula assumption when it estimates portfolio VaR from individual asset VaRs. Today, with SPY -1.40%, QQQ -1.39%, IWM -1.61%, HYG -0.51%, and GLD -3.16%, you can see the correlation structure collapsing in real time.
Computing Portfolio ES Under a Copula Model
[Python: portfolio Expected Shortfall from copula-simulated return matrix — 99% confidence, SPY/QQQ equal weight, Gaussian vs Clayton comparison showing ~20-30% gap]
For a $100,000 SPY/QQQ portfolio, the Gaussian model might give you a 99% daily ES of $3,200. The Clayton model says $4,000. That $800 gap is not a rounding error. It's the hidden tail exposure that doesn't show up in your risk report because the model was chosen for convenience, not accuracy.
Trade Implications
Today's numbers make this concrete. VIX at 25.09 prices annualized vol at 25% in SPY options. The 99% daily ES under a Student-t(5) model with that vol is roughly 5.8%, about $38 per SPY share. For a 100-share position ($66,143 notional), your expected conditional tail loss on your worst 1% of days is around $3,800. Not $2,400 as the VaR suggests. $3,800.
Under Clayton copula for a correlated portfolio, portfolio ES is higher than the weighted average of individual ES values. Diversification does not help in this regime. GLD being down 3.16% on the same day as equities confirms the correlation breakdown is already happening.
Three things follow.
Position sizing: if you're scaling to a normal-distribution VaR target, you're likely running 1.5x to 2x more tail exposure than your numbers say. Scale notional by the ES/VaR ratio from your Student-t fit, not the normal-model estimate of 1.13.
Volatility adjustment: VIX at 25 vs. a 12-month average of 17 to 18 represents a 40% increase in daily vol. If you didn't trim exposure when VIX crossed 20, you're holding 40% more risk than your baseline model assumed.
Hedging cost: a 5% OTM SPY put expiring in 30 days costs substantially more with VIX at 25 than at 15. Whether that premium is worth paying depends on your actual tail distribution, not the Black-Scholes normal assumption. Under a fat-tailed, Clayton-copula model, the answer tilts toward yes. The vol risk premium from short vol strategies compresses or vanishes when VIX is above 25, precisely when the tail protection you'd be selling costs the most to replace.
Getting the Copula Choice Right
The practical cost of using a Gaussian copula in a correlated equity portfolio is a 15 to 25% systematic underestimate of true tail exposure. You discover this through backtesting: count how often your portfolio loss exceeds the 99% VaR estimate. Under a correct model, that happens 1% of the time. In practice, for equity portfolios with Gaussian copula assumptions, it happens 2 to 4% of the time. Your 100-year flood arrives every 25 years.
Switching to a t-copula or Clayton copula won't prevent blowups. It will at least make your risk estimates honest. You hold less gross exposure. You pay for more protection. You survive the tail events that your Gaussian model classified as once-in-30-years events but that arrive with boring regularity.
Days like today, with VIX up 12% in a session, are not 30-year events. They're 6-month events. Your model needs to reflect that, not the smooth interior of a normal distribution where it's comfortable and wrong.
The tail is not an edge case. It's a schedule.
#trading #quant #algotrading #quantfinance
109
Due Diligence: SoFi Technologies (SOFI)
Current price: $17.37. Market cap: $22.2B. Rating: HOLD with bearish tilt.
GAAP net income: $406M. Subtract $213M SBC and economic earnings drop to $193M. Economic P/E: 111-148x, not 55x.
Balance sheet grew from $7.3B to $34.9B since 2021. Cost: 50% shareholder dilution. Permanent.
Muddy Waters: 6.1% charge-off rate vs 2.89% reported. $312M potential unrecorded debt. Loan Platform Business may be disguised borrowing.
Debate: Bear 7 wins. Bull 1 win. 3 draws.
Probability scenarios:
Bull 20%: $29
Base 50%: $18.50
Bear 30%: $10
Key unknown: NCO rates on $28.5B personal loan book in FY2025 10-K.
Levels: Support $16. Exit below $14. Sell if NCOs > 4.5%.
Full research: tradescript.substack.com/p/s…
#SOFI #quant #fintwit Due Diligence: SoFi Technologies (SOFI)
20










