Summer 2026 · BSc Economics, University of Surrey

A Structured Value Research Process for UK Small Caps

And the parts of it that didn't work.

A repeatable research process for finding mispriced companies in a part of the market large institutions cannot easily reach: AIM and the lower end of the FTSE SmallCap. 785 listed companies, narrowed to five detailed company studies, with every step reproducible from source files — and a written record of where the process went wrong.

785
Raw universe, 13 Aug 2026
410
Matched screening universe
57
Reading queue after five gates
5
Rebuilt from primary filings
Three things this project demonstrates
01

A reproducible screening pipeline

Built on data nobody sells as a single product. A tradability gate, a data-availability triage, and five deterministic gates. The gates are arithmetic, not judgement, and they are versioned.

The design →
02

Primary-document research, done properly

Five companies rebuilt from five to seven years of annual reports into full three-statement models. Each valued in the frame appropriate to it, and each built without looking at the share price until the range was finished.

Five companies →
03

An unexpected finding about the tool itself

I began the summer thinking the interesting question about AI in research was what it can do. By August I thought the interesting question was how it fails, specifically, in this workflow.

AI as a research tool →
The thing I did not expect to learn

Numbers from a database, reasoning from the model, judgment and decision from the human.

Coherence bias operates on synthesis and framing. It does not operate on checkable facts. That gives a clean division of labour rather than a vague warning — and it is the empirical justification for a design rule I had already written down without evidence.

Read section 08

The funnel

Every stage regenerates from the published CSVs. Fifty-two names in the final queue have not been looked at — that is the honest current state.

Funnel diagram: 785 narrowed to 5
785 → 5 · band width scales with count
StageCountWhat happened
Raw universe785AIM full list (597) plus FTSE SmallCap constituents (188)
IBKR tradable74441 dropped as untradeable through my own broker
Triage275 / 196 / 314Clean pass / held on a data gap / excluded
Held names promoted+135Volume confirmed through a second source
Screening universe410Matched to fundamentals data
Five deterministic gates5738 AIM, 19 FTSE SmallCap
Examined in detail5ARBB, KMK, NAH, SNWS, TFW

Contents

The sections run in the order the work happened. Every section follows the same shape: the claim, the test, the result, and what it changed about the design. Sections 08 and 10 are the ones I would read first if I only had five minutes.

01
Where this started, and why it collapsed →
Nine thousand backtests, and why they didn't survive
02
From cognitive edge to structural edge →
Where institutional capital cannot go
03
Choosing where to look →
Why the UK, and not the cheaper markets
04
The design →
Four premises, seven components, and the anti-scope
05
The data problem →
Every source that covers AIM fails on access
06
The pipeline: 785 to 57 →
Five deterministic gates, and the model I threw away
07
Five companies, read properly →
ARBB, KMK, NAH, SNWS, TFW
08
AI as a research tool, and how it fails →
The three-layer control framework
09
Method and tooling →
Four utilities, and what is still manual
10
What went wrong, and what happens next →
Four errors, and the next four pieces of work
11
About →
Who wrote this, and why
Sections 08 and 10 are the ones I would read first if I only had five minutes.