correl-range

correlated variable monte carlo simulations

• Example • API • Notes • License

Example

import SIM from '../sim.js'

const stats = SIM(
({N,L,W,U})=>{
// initiation ran once
const fixed$ = L(500_000, 650_000, 'demand', 0.5, 'price'),
            month$ = L(5_000, 7_000, 'demand', 0.5, 'season', 0.5),
            months = L(6, 9, 'season', 0.5, 'price', -0.5)
return ()=>{
// calculations on every iterations
const total$ = fixed$ + month$ * months
return {
// exported results
    months,
    total$
}}}
).run(10_000).stats

console.log('total$ range', stats.total$.Q(0.25).toFixed(0), stats.total$.Q(0.75).toFixed(0))
console.log('correlation', stats.total$.cor('months'))

API

sim( factory, {confidence=0.5, resolution=128} ).run( N=25_000 ) ⇒ simulation

• factory: ({N, L, W, U}) => model where N, L, W and D are randomVariableFactory
• randomVariableFactory: (low, high, ...correlation) => randomVariable to match the simulation confidence interval
  • N: normal
  • L: lognormal
  • W: weibull
  • U: uniform
• randomVariable: with .valueOf() that changes on each iteration
• simulation
  • stats: empirical distribution cdf, pdf, quantiles, average (based on modules sample-distribution and lazy-stats)

Notes

1. use case is human approximation in decision making - "guesstimates"
2. default is to use a confidence interval of 50% (IQR)

• familiarity with box plots
• minimizes overconfidence
• 50% is also conservative approximation of the median min and max of 3 occurences (2^(1-1/n)-1=59% for n=3). This is in line with studies showing ~50% confidence range when asked to provide ~60% confidence. Actual overconfidence varies between studies but is always present.

3. variables can be correlated with independent risk factors by providing the linear factor
4. to maintain correlation, each variable returns a single value per cycle - random variables are constant within a given cycle

License

MIT © Hugo Villeneuve