Adjacent-Signal Nowcasting
Infer an unobservable current state from multiple related measurements
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 5
- Confidence
- 96%
Nowcasting estimates an important present quantity that cannot be observed directly. The practitioner first defines the hidden state, such as the current fair value of a thinly traded bond. They then gather adjacent evidence: recent transactions, dealer quotes, issuer information, related instruments, or other measurements whose relationship to the hidden quantity can be modeled. Statistical estimation fuses these imperfect and differently timed observations into a best current guess, ideally with an uncertainty range. This is distinct from forecasting because the target exists now; the problem is incomplete observation rather than an unknown future. McAuliffe compares it to aircraft control, where software estimates position and velocity from GPS and instruments because nature does not directly provide perfectly exact state variables. Once the current state is estimated, prediction, portfolio construction, or control can proceed on firmer ground.
Origin
Extracted from Masters in Business through Jon McAuliffe's explanation of corporate-bond fair-value estimation and automated aircraft control.
Core principles
- 01A crucial current state may be unobservable even when related evidence is abundant
- 02Combine multiple imperfect measurements rather than relying on one stale proxy
- 03Model the present before attempting to forecast the future
- 04Domain structure determines which adjacent signals are informative
How to run it
- 1
Name the hidden present state
Specify the current quantity required for a decision but unavailable as a direct, reliable observation.
Pro tip Distinguish state estimation from future forecasting.
Watch out An ambiguous target produces an incoherent collection of proxies.
- 2
Gather adjacent measurements
Collect recent trades, quotes, related assets, fundamentals, sensors, and other observations connected to the hidden state.
Pro tip Record timestamps and provenance for every measurement.
Watch out Stale or one-sided observations should not be treated as current truth.
- 3
Model reliability and relationships
Estimate how strongly each measurement relates to the target and how its error changes with age or conditions.
Pro tip Allow source reliability to vary over time.
Watch out Equal weighting can let weak evidence dominate stronger signals.
- 4
Fuse the evidence
Combine the available measurements into a present-state estimate and quantify uncertainty where possible.
Pro tip Use domain constraints to reject impossible estimates.
Watch out A point estimate without uncertainty may encourage excessive confidence.
- 5
Back-check and update
When direct observations eventually arrive, compare them with prior nowcasts and recalibrate the model.
Pro tip Track errors by liquidity regime and source mix.
Watch out Structural changes can make historical relationships obsolete.
In the wild
A bond has not traded today and lacks a liquid live order book. The estimator combines yesterday's transactions, recent one-sided dealer quotes, issuer information, and related market movements to infer its current fair value before a trading decision is made.
→ The trading system gains a defensible present-price estimate instead of relying on a single stale print.
Flight-control software needs current position and velocity, but no instrument supplies all six values with perfect precision. It combines GPS and other onboard measurements to estimate the aircraft's present state continuously.
→ The controller receives usable state estimates for safe automated adjustments.
Common mistakes
Forecasting before locating
Trying to predict change before estimating the current state compounds uncertainty and can invalidate the forecast target.
Trusting the latest print
The most recent observation may still be stale, unrepresentative, or generated under different market conditions.
Hiding estimation uncertainty
Treating an inferred value as exact causes downstream systems to take more risk than the evidence supports.
Is it for you?
Best for
It is best for illiquid markets, sensor systems, and operational settings with partial or noisy observations.
Not ideal for
It is not ideal when an accurate direct measurement is already cheaply and continuously available.
From the transcript
“I wish I knew what the fair value of this of this CSIP was.”
“I have lots of stuff that's related to the number I know that I want to know. I just don't know that number”
“you can use statistics to go from all that stuff that's adjacent to a guess and infill of the thing you wish you knew”
From the episode
Jon McAuliffe on Innovation and Statistical Methods