A learner that knows when to say nothing

Summary

Across 10 seeds per scenario, the ARCANE structural learner beat the base rate wherever structure had been planted, and in none of 30 seeded runs did it validate a rule built on noise. On 515 walk-forward predictions of the US high-yield credit spread it found nothing that survived validation, abstained throughout, and scored exactly what climatology scored.

Results

Brier score, learner vs base rate (lower is better)

Series	Learner	Base rate	Difference [95% CI]
Synthetic: lagged chain (10 seeds)	0.2204	0.2488	−0.0285 [−0.0339, −0.0231]
Synthetic: regime switch (10 seeds)	0.2094	0.2362	−0.0268 [−0.0315, −0.0222]
Synthetic: no structure (10 seeds)	0.2300	0.2275	+0.0025 [+0.0019, +0.0031]
US high-yield OAS, 515 predictions	0.2434	0.2434	0.0000

Synthetic rows use a rolling-window base rate and a percentile bootstrap over seeds (4,000 resamples). The credit row uses climatology and a moving-block bootstrap.

The question

Any large enough system contains patterns, including a random one. So the useful property of a pattern-finder is not that it finds things. It is that it refuses the things that are not there, and says so.

Setup

Three synthetic worlds of 1,200 steps each: one with a lagged causal chain, one whose rule changes at step 600, and one with no structure at all. The learner proposed rules every 25 steps from step 150 and was scored from step 300. Each world ran on 10 seeds.

The real test used the ICE BofA US High Yield option-adjusted spread from FRED: 786 daily observations from 18 September 2023 to 15 September 2026. The target was whether the next observation would be strictly higher. After a 250-day warm-up, that gave 515 walk-forward predictions.

What we found

The planted single-condition rule was recovered in 10 of 10 seeds in both structured worlds. A harder two-condition rule was recovered in 8 of 10. When the regime switched, the old rule was retired in all 10 seeds, after 84 steps on average, and a replacement was validated in all 10, after 200 steps on average.

On the world with no structure, no noise rule was validated in any seed. The learner spoke on about 3% of steps and paid a small cost for it, +0.0025 Brier against the base rate.

On the credit spread, 52 discovery rounds produced no rule that survived validation. The learner abstained on every prediction and matched climatology to four decimal places. The strongest rejected candidate had negative out-of-sample skill.

Why we publish a zero

A learner that had found a rule on the credit spread would have made a better headline. It would also have been harder to believe. The synthetic runs show the learner can find real structure; the credit run shows it does not manufacture any when it cannot.

Limitations

The synthetic worlds were designed by us, so recovering their structure is a necessary test, not a sufficient one.
One real series and one target. A zero on the high-yield spread says nothing about other series.
A planned comparison with an external learning system did not produce a valid run and is not reported.

What would change this

A rule validated on a fresh seed set built with no structure.
A rule on a real series that validates and then holds on data it has never seen.

Source boundary

Published: scenarios, protocol, scores, intervals and recovery counts. Withheld: the rule grammar, validation thresholds and penalties, and feature construction.

Changelog

17 Sep 2026Synthetic and credit-spread experiments run.
22 Sep 2026First public edition on Labs.

Cite

@techreport{arcane2026learnerthat,
  title       = {A learner that knows when to say nothing},
  author      = {{ARCANE Labs}},
  institution = {ARCANE Intel},
  type        = {Paper},
  year        = {2026},
  url         = {https://arcaneintel.net/labs/learner-that-abstains}
}