Methodology
About the Dojo Score
A single 0–10 score that ranks Japan's small & mid caps on the evidence in their own disclosures — built with strict research discipline and reported honestly.
What goes into the score
The Dojo Score blends a small set of well-understood, point-in-time factors plus one signal that's distinctive to our coverage:
- Value — valuation relative to the company's own fundamentals.
- Quality — profitability and balance-sheet health.
- Momentum — medium-term trend, de-emphasised (it was weak on its own — see below).
- Size — the small/mid-cap tilt where English coverage is thinnest.
- 5% large-shareholder flow — net accumulation/withdrawal from statutory large-holding filings; the differentiator, kept even though its standalone signal is modest, because it aligns with what we cover.
The output is mapped to a fixed 0–10 band so a score means the same thing across dates — it is not re-scaled to the latest cross-section.
How we test it (and don't fool ourselves)
- Point-in-time, strict. Every factor at a date uses only data that was public by then — no lookahead.
- Walk-forward. The model is trained on a development window and evaluated going forward.
- Net of costs. Returns are after estimated trading costs, not gross.
- One sealed hold-out. A 24-month hold-out (Mar 2024 – Feb 2026) was locked away during development and tested exactly once. We don't re-run it.
Published results — v1 sealed hold-out
These are the only performance figures we cite. The development column is in-sample (the model learned on it); the sealed hold-out column is the one-time forward test. All figures are backtested and net of estimated costs. Benchmarks are described generically — see the note.
The same out-of-sample curve we show on the home page, in full — our locked v1, indexed to 100, after estimated costs. The shaded band is the sealed hold-out (Mar 2024 – Feb 2026) we tested exactly once; the dotted line is the S&P 500 for global context:
| Metric | Development (in-sample) | Sealed hold-out (tested once) |
|---|---|---|
| Rank IC (NW t-stat) | +0.129 (t 4.8) | +0.144 (t 3.4) |
| IC win rate | 90% | 92% |
| Decile monotonicity (Spearman) | 0.99 | 0.98 |
| Top-decile minus bottom (D10−D1) / 3M | +4.8% (t 3.4) | +7.1% (t 2.7) |
| Top-decile excess / 3M | +1.7% (t 3.0) | +4.0% (t 2.7) |
| Annual excess vs small-cap benchmark | +6.1%/yr | +3.9%/yr (t 1.2) |
Our read, plainly: the hold-out excess (+3.9%/yr vs a broad small-cap benchmark) clears our pre-registered +3%/yr bar and the ranking stays well-ordered, but the 24-month t-stat (≈1.2) is not high enough to call it conclusive. We describe v1 as promising but not yet proven over this window — not a guarantee.
What didn't work (we publish this too)
- News/text (NLP) features — added noise, not signal: development Rank IC got worse (0.129 → 0.121). Dropped, and recorded as an honest negative result.
- Concentrated Top-N portfolios — net returns were negative after costs. Not used.
- Plain deciles with no smoothing — net −1.0%. Not used; we smooth instead.
What you won't find here
No price charts, no ¥ prices, no PER/PBR or other raw vendor values — only our own derived analytics (excess returns, deciles, information coefficients). Raw Japanese price and index data carry licensing we haven't cleared for redistribution, so we don't display them.