Quincus

Estimation and ground truth

The model is public. The calibration is not.

Every method on this site is published somewhere. Column generation is in textbooks, jump diffusion is in every quantitative finance course, and attention models for routing are on arXiv. What is not published is what the parameters actually are on a specific lane, in a specific season, under a specific carrier. Quincus has eleven years of being the system that made the call and then found out whether it was right.

LIVE

Calibrate a lane

Estimator

Truth is lognormal with log-scale volatility above. Log units on the x axis.

0.000.250.500.751.000.000.250.500.751.00Predicted probabilityObserved frequencyPerfect calibrationCalibrated

The calibrated estimator is at its structural floor. Adding observations tightens the fit but the underlying volatility keeps the interval this wide.

Interval width
8.36d
90% predictive interval, held-out. Narrows with observations, then flattens at the structural floor set by true volatility.
Observations200
CRPS1.336
Coverage @ 90%85.0%
Interval width8.36d
Calibration error0.061
Estimatorcalibrated
Runs in your browser on a simplified version of the production method. Illustrative of behavior, not of production performance.

The problem class

Three things only operating gives you.

01

Predicted versus realized.

Every arrival time the platform issued, held against what actually happened, conditioned on lane, season, carrier, and regime.

02

Quoted versus won.

Every price offered and whether it converted. The conversion curve is not derivable from public data at any price.

03

Described versus found.

The mapping from an informal address to a physical delivery point, in networks where formal addressing does not exist.

The method

How estimation is done here.

Calibration, not fitting
θ^=argmaxθt=1Tlogf ⁣(ytFt1;θ)\hat{\theta} = \arg\max_{\theta} \sum_{t=1}^{T} \log f\!\left( y_t \mid \mathcal{F}_{t-1}; \theta \right)

Regime transition probabilities, jump intensities, and duration distributions are estimated against realized outcomes rather than assumed. The functional form is standard. The parameter values are the asset.

Distributions, scored
CRPS(F,y)=(F(x)1{xy})2dx\mathrm{CRPS}(F, y) = \int_{-\infty}^{\infty} \big( F(x) - \mathbf{1}\{x \ge y\} \big)^2 \, dx

A point estimate can only be wrong by an amount. A distribution can be wrong in shape, and the continuous ranked probability score measures both calibration and sharpness at once. Every estimator in the platform is scored this way against realized outcomes, continuously.

Calibration as a testable property
Pr ⁣(YF^1(q))=qq(0,1)\Pr\!\left( Y \le \hat{F}^{-1}(q) \right) = q \quad \forall q \in (0,1)

A ninety percent interval that contains the outcome seventy percent of the time is not conservative, it is wrong. Calibration is checked per lane and per regime, so drift is detected as it develops rather than discovered in a quarterly review.

Priors that inherit
p(θD)p(Dθ)p(θ)p(\theta \mid \mathcal{D}) \propto p(\mathcal{D} \mid \theta) \, p(\theta)

A new lane does not start blind. It inherits a prior from structurally similar lanes and updates toward its own behavior as observations accumulate, which is why coverage on a new corridor is useful in weeks rather than years.

Why this is hard to replicate

You cannot scrape being wrong.

A foundation model trained on the public internet has never dispatched a shipment and never found out whether it arrived. This class of data is generated only by sitting in the decision loop and bearing the consequence. It cannot be scraped, licensed, or synthesized, because it does not exist anywhere until someone commits to a decision and the world responds.

The methods on this site are reproducible. A competent team with a good model could implement column generation in weeks. What that team would not have is what lambda is on a specific corridor in the third quarter, what a quote at a given price actually converts at for a given customer, or where a delivery point physically sits when the address is a description rather than a number. Those are earned in production, and the only way to acquire them is to be wrong in public, at a customer's expense, for a long time.

11
years of predicted versus realized
7
countries of location ground truth
1
decision loop, closed the entire time

What this changes

What calibration buys.

01

Every optimization downstream consumes a distribution that has been scored against outcomes rather than a vendor's point estimate.

02

Model drift surfaces as a calibration failure on a specific lane rather than as an unexplained margin miss.

03

A new entrant can reproduce the method in a month and still be years away from the parameters.

If your network makes decisions under uncertainty, we should talk.

We work with a small number of operators at a time. Tell us what your network is optimizing for.

Start a conversation