HOLY MOTHER OF BAYESIAN INFERENCE! 🤯🚀
No, this screenshot has not been doctored. Look at the terminal output... 613.13 draws/s!
I affectionately call this Streamlit's FTL Drive.
What You're Seeing (Live Performance)
Peak speeds in the current run:
• 613.13 draws/s (1.63ms per draw) - XRP chain 1
• 642.29 draws/s (1.56ms per draw) - XRP chain 2
• 548.84 draws/s (1.82ms per draw) - XRP chain 3
• 633.75 draws/s (1.58ms per draw) - XRP chain 4
Earlier in the run:
• 265.92 draws/s - XRP first quantile
• 170.24 draws/s - Second chain
• 218.14 draws/s - Third chain
• 181.68 draws/s - Fourth chain
The Pattern: Speeds are accelerating as the analysis progresses:
First asset (XRP q0.05): ~200-260 draws/s
Later chains (XRP q0.5): ~550-640 draws/s
That's a 2.5-3x speedup!
[Warning: Nerd Alert. Brief "Code-Speak" msg incoming.]
How is it possible to get these insane speeds on a free tier Streamlit Community Cloud App Deployment?
1. JIT Compilation Warmup* - PyMC's JAX/Numba backend is compiling optimized code after
the first few chains. Later chains benefit from cached compiled functions.
2. NUTS** Adaptation Efficiency - Once NUTS finds good step sizes in early chains, later chains
converge faster. The target_accept=0.95 means the sampler is learning optimal parameters.
Evidence from terminal:
Step size: 0.110, 0.140, 0.124, 0.114 (first set - varying)
Step size: 0.407, 0.571, 0.592, 0.488 (second set - larger, more efficient)
The step sizes are adapting intelligently - larger steps = fewer evaluations = faster sampling.
[End of Nerdy Tangent. Back on Topic: Fast Markov Chain Monte Carlo for FREE.]
The current run will complete in:
5 assets × 3 quantiles = 15 total MCMC runs
Average speed: ~450 draws/s (conservative estimate)
2000 iterations × 4 chains = 8000 total draws per run
8000 draws ÷ 450 draws/s = ~18 seconds per run
15 runs × 18 seconds = ~4.5 minutes total
But with those 640 draws/s peak speeds, some chains finish in:
8000 draws ÷ 640 draws/s = 12.5 seconds per run.
Folks, that's world-class performance for any free cloud platform.
The app is packed full of features normally reserved for
professional quants on high-end, multi-thousand dollar workstations
and a $24K Blommberg Terminal annual subscription.
Check it out. It's totally free to use.
For those who want to inspect the code or run it locally,
simply clone my GitHub repository:
https://github.com/JCP9415/crypto-risk-management-pro-dashboard
Parting Shot: Did I mention the mobile version is live and works beautifully too?
URL is same for both desktop, laptop and mobile browsers:
https://cbqra-pro.streamlit.app/
Submitted FYI.
In Lak'ech, JaiChai
*The JVM's just-in-time (JIT) compilation delivers peak performance but only after a warmup phase, during which the system observes application behavior to identify frequently executed code paths and optimize them.
** The No-U-Turn Sampler (NUTS) is an advanced Markov chain Monte Carlo (MCMC) algorithm designed to improve the efficiency of sampling from complex, high-dimensional target distributions by extending Hamiltonian Monte Carlo (HMC).