Letter to the Judiciary of Pakistan by Arbitrarily detained Former Prime Minister Imran Khan, transcribed by his lawyer Barrister Salman Safdar, based on the recent conversation in Adiala Jail on November 11, 2026.
اہم اعلان : میرے خیال میں مارچ کے سینیٹ انتخابات میں پنجاب سے احد خان صاحب کو پی ٹی آئی کا سینیٹر بنانا چاہیے۔ جس دلیری اور بے باکی سے وہ بات کر رہے ہیں، ہم جماعت والوں میں سے کوئی ویسے نہیں بول پایا۔
پی ٹی آئی کو اس وقت روایتی سیاست کی نہیں، دلیر، وفادار، باکردار اور بلا خوف ظالم کو للکارنے والی آوازوں کی ضرورت ہے۔ آئندہ ٹکٹ بھی صرف ایسے امیدواروں کو ملنے چاہییں جو اس معیار پر پورا اتریں۔ میری نظر میں احد خان صاحب سے بہتر انتخاب پی ٹی آئی کے پاس کوئی نہیں۔
ایک بات پہلے ہی واضح کر دوں کہ مجھے موروثیت کا لیکچر نہ دیجیے گا۔ میں خود موروثیت کے خلاف ہوں۔ میں احد خان کو موروثیت کی بنیاد پر نہیں، ان کی دلیری، نظریاتی وضاحت اور بغیر کسی اگر مگر کے کھل کر بات کرنے کی جرأت کی بنیاد پر نامزد کر رہا ہوں۔ خان کا رشتہ دار ہونا کوئی جرم نہیں۔ اصل سوال یہ ہے کہ نظریے کے ساتھ حقیقت میں کون کھڑا ہے۔ اور پھر پی ٹی آئی کے تقریباً ہر بڑے رہنما کے گھر سے کوئی نہ کوئی اسمبلی میں بیٹھا ہے، تو کیا یہ قدغن صرف خان کے گھرانے کے لیے ہے؟
نوٹ: مجھے معلوم ہے کہ احد خان صاحب نہ اس کے خواہش مند ہیں اور نہ ہی شاید اسے قبول کریں گے۔ لیکن ہمیں خان صاحب کی آزادی کی خاطر ان سے درخواست کرنی چاہیے کہ وہ سینیٹ میں آئیں۔
I did it. Officially a CQF graduate 🎓 Late nights, broken code, plenty of self-doubt, and I made it through. Built a neural network to price options on 100 assets. Next chapter: CFA 📚 Let's connect and learn together: https://t.co/EBWYwb2TPC
#CQF#QuantFinance@CQFInstitute
Black-Scholes has one input that never moves, and the market disagrees with it every single day.
The smile exists. Equity skew leans left. A single constant σ cannot produce either shape, no matter how carefully you pick it.
Heston's 1993 answer was to stop treating volatility as a number and start treating it as a state variable:
• the stock diffuses with √v_t
• the variance follows a Cox-Ingersoll-Ross square-root process that mean-reverts to θ at speed κ
• the two Brownian shocks are correlated by ρ
That last parameter is where the model stops describing the skew and starts explaining it. With ρ < 0, a down move in the stock arrives with an up move in variance. Downside protection gets mechanically more expensive, and out-of-the-money puts trade at richer implied vols than equidistant calls.
Two details from building this out in Python:
1. The Feller condition (2κθ > ξ²) is violated by the demo parameters — 0.16 < 0.25. That is not a bug, it is typical of real calibrations. It just means zero is attainable and your simulation scheme has to respect the boundary. Full truncation, v⁺ = max(v, 0), is the usual first response.
2. The ATM implied vol from the model was 19.21%, not √θ = 20%. Under stochastic volatility the mapping from parameters to implied vol is indirect. The market never observes θ. It observes prices.
The post derives the SDE, builds a NumPy Monte Carlo pricer from scratch, then cross-checks it against three QuantLib engines on the same contract:
• Analytic (characteristic function): 10.1546
• Finite difference: 10.1542
• QuantLib MC: 10.1751
• Scratch NumPy MC: 10.1771
Which do you reach for first when a model has both a semi-analytic transform and a PDE grid available?
https://t.co/WZYbKrIiav
#QuantFinance #OptionPricing #StochasticVolatility #Python #FinancialEngineering
Black-Scholes doesn't break where most people think it does. 📉
It's not log-normality or continuous trading that fails first. It's the assumption that one constant volatility number can price every strike and every expiry on the sheet.
Real markets never quote a single vol. They quote a whole map — and that map is the actual signal.
Here's what the post unpacks, from first principles to Python:
→ Implied vol isn't a property of the stock. It's the σ you back out by inverting an option price. Same stock, different strikes and expiries → different implied vols. That's not a contradiction; it's information.
→ The smile is systematic, not noise. Price a 9-strike equity chain at a flat 20.5% ATM vol and you underprice downside puts every time. An 85-strike 6-month put jumps from 0.64 (flat) to 1.04 (surface) — 63% richer, exactly where risk managers care most.
→ Two axes, not one. Strike gives you the skew; expiry gives you the term structure. Compress that 2D surface into a single scalar and you're not approximating — you're losing information.
The post builds all of it in code: brentq inversion, a SABR-generated surface in QuantLib, a numerical Dupire local-vol reconstruction, and a flat-vol vs. surface-vol pricing comparison.
My take: stop treating volatility as a nuisance parameter to calibrate away. It's the market telling you precisely how it prices crash risk, leverage, and jump fear. Volatility is the signal.
Which volatility model do you actually reach for in production — local vol, Heston, or something jumpier? 👇
Full breakdown with code and charts:
https://t.co/fgN0hr4gYJ
#QuantFinance #OptionsPricing #Volatility #Python #QuantLib
Every market-regime tutorial ends with a beautiful green/amber/red chart.
Almost none answer the expensive question: what decision should actually change because of that label?
So I tested it instead of hand-waving.
The setup: nine US sector ETFs, 1,283 weekly feature–target observations, six common features, three VIX states, three walk-forward policies. After a five-year warm-up, the test covered 1,023 decisions:
→ one unconditional feature model
→ hard regime gating
→ soft regime-conditional weighting
The result was a useful failure.
None of the apparent low-vs-high volatility differences survived block-bootstrap uncertainty and multiple-testing correction. And the regime layer made execution worse:
→ Unconditional turnover: 59.4×/yr
→ Hard gate: 73.6×
→ Soft conditional: 92.6×
At 10 bps per unit of turnover, the soft policy lost 12.46% annualized vs 8.19% for the already-weak baseline.
Then I decomposed that 4.58-point shortfall — and found the part that matters:
→ Return effect: −1.26 pts/yr, |t| = 0.40 → indistinguishable from zero
→ Turnover bill: −3.32 pts/yr, |t| = 10.73 → effectively a certainty
The same 19.7 years measured the cost of the regime layer with ~27× more statistical resolution than its benefit.
That's the uncomfortable general lesson: a backtest has plenty of power to measure what a regime layer costs, and almost none to measure what it earns.
So the question before shipping one isn't "does the backtest look better?"
It's: which half of that comparison did my data actually have the power to measure?
Regime detection isn't about naming the market. Its real value is deciding when not to trust an otherwise useful relationship — and proving that decision survives uncertainty, walk-forward validation, and costs.
Full write-up + downloadable notebook (with deterministic offline mode) 👇
What downstream decision does your regime model actually improve?
https://t.co/NqFtWNhI2m
#QuantFinance #MachineLearning #DataScience #AlgorithmicTrading #Backtesting
In the late 1980s, Dennis Weatherstone at J.P. Morgan asked a deceptively simple question: "How much money could we lose tomorrow?"
That question eventually gave us Value at Risk — the risk metric that became the shared language of every trading desk, regulator, and CFO on the planet. Then 2008 happened, and everyone realized what VaR had been hiding all along.
My latest post in the Randomness to Risk series builds four VaR models from scratch in Python and then stress-tests them against a regime-switching return series with fat tails:
Historical Simulation — no distributional assumptions, but painfully slow to react when volatility spikes
Rolling Parametric (Normal) — analytically clean, fast, and dangerously wrong when markets go non-normal
EWMA (the original RiskMetrics methodology, λ=0.94) — exponentially weighted, 11-day half-life on information, reacts to crises faster than the rolling window alternatives
Student-t VaR — accounts for fat tails where a 4-sigma move that theory says happens once in 31,574 days actually shows up six times in the data
Then I run the Kupiec Proportion of Failures backtest on all of them, implement the Basel traffic light framework (green / yellow / red exception zones), and show exactly where each model breaks.
The honest result? All three rolling models get rejected on unconditional coverage. EWMA is the least bad. None of them pass cleanly.
The takeaway is not that VaR is useless. It is that VaR is a starting point, not a destination. Expected Shortfall, stress testing, and regime awareness are what you add on top.
Full post with working code here: https://t.co/ylOD5WG3dQ
What risk model do you rely on most in practice — and have you ever seen it fail catastrophically at the worst possible moment?
#QuantitativeFinance #RiskManagement #Python #AlgoTrading #DataScience
The Complete Option Pricing Series
Option pricing looks like a collection of formulas—until you follow the full chain of ideas.
Across June and July, I built an eight-part path from modeling random stock prices to pricing contracts with path dependence and early exercise. The progression made one lesson especially clear: there is no universally “best” pricing method. The right tool depends on the payoff, dimensionality, accuracy requirements, and whether exercise is a decision.
Here is the complete reading path:
1. From Itô to Black-Scholes — how stochastic calculus turned randomness into a pricing engine
https://t.co/Z5Xji4C1ni
2. Geometric Brownian Motion in Python — build and challenge the first stock-price model most quants learn
https://t.co/1OIZlSKwsf
3. Black-Scholes Without the Magic — derive the pricing PDE through delta hedging and no-arbitrage
https://t.co/E5A33lsyHi
4. Monte Carlo Option Pricing — move beyond closed forms to Asian, barrier, and lookback payoffs
https://t.co/SNSBvX8ole
5. The American Option Problem — understand optimal stopping, the early-exercise premium, and the free boundary
https://t.co/9VD9RxpBNW
6. Binomial Trees for American Puts — compare CRR and Leisen-Reimer convergence through backward induction
https://t.co/u1GXJ6oWcj
7. Finite Differences and PSOR — solve the American put PDE as a linear complementarity problem
https://t.co/7BZh7ymu1z
8. Longstaff-Schwartz — combine forward simulation with backward regression to estimate the exercise policy
https://t.co/nS2L6iDMZF
Together, these articles connect the theory to reproducible Python and show why closed forms, simulation, trees, and PDE grids are complementary rather than competing ideas.
Which pricing method would you reach for first—and what would make you choose a different one?
#QuantFinance #OptionPricing #FinancialEngineering #MonteCarlo #Python
1. Elite Capture. When you are going meet your fellow politicians next time, just look around and access their pedigree.
2. Elite Capture. People came out many times and faced live fire power of the motherland. Their loved ones still waiting for dead bodies. Not a single elite even got little rash.
3. Elite Capture. You all party together and want people to .… I would say learn from Dr. Abdul Qadeer Khan’s fate.
Pakistan = Vertical Integration Criminal Enterprise
Innovation thrives on openness, not restrictions. America wins the AI race by empowering builders, not by locking down code. Fully agree—let's make American Open-Source AI the global standard. 🇺🇸🚀
We need to incentivize American Opensource AI not ban Opensource AI
Both @MikeBradleyAI and I started @OsmanticAI because we believe that we need to make it the default
The current administration needs to work with people like us to make American Opensource AI THE BEST there is