A-share micro-cap history

Historical association · 2008–2026 · 223 month ends

Industry-demeaned factor spreads vary across regimes. Size is defined in the large-cap direction; a negative spread cannot be described as small-cap underperformance.

The claim that 2026 was the first positive large-cap year conflicts with positive years listed in the same record. Early fundamentals are reconstructed; 2026 is incomplete.

Exploratory comparison · Overlapping starts in 2019–2022, ending in 2026

Selected long-short combinations appeared stable across several starts. These are overlapping histories rather than independent tests.

A factor that helped a long-short spread need not help a long-only account. Keep selected model weights restricted and require a frozen future comparison.

Endpoint-sensitive · Historical windows ending in 2026

An attractive minimum-cap filter depended on the ending period. Changing the endpoint weakened the original interpretation.

Neither a permanent size floor nor unrestricted micro-cap exposure is established as the long-run optimum.

Historical simulation · 2008–2026; separate post-2015 universe

Long-short findings did not transfer mechanically to long-only portfolios. Results changed materially with the universe and starting date.

Early unadjusted returns and incomplete eligibility require repair. A broad-universe result cannot certify a more restricted investable universe.

Selection-sensitive search · 455 combinations of 14 factors

The historical search found attractive selected combinations, with substantial drawdowns and weaker later-period results.

The selected winner is not independent proof of optimal synergy. Conflicting annualizations and an inconsistent yearly win count remain unresolved.

Construction sensitivity · 2016–August 2026

A combined-score screen exceeded separate sleeves in the inspected history; much of the advantage appeared after 2021.

Full-sample selection prevents a strict out-of-sample claim. Match the eligibility, cash, costs and execution clock before selecting a construction.

Corrected assumptions · Separate 2019–2026 execution checks

Capital assumptions and a 1,000-fold traded-amount unit error changed earlier liquidity interpretations. A high simulated fill ratio followed the correction.

A high fill ratio under an assumed model does not establish unconstrained liquidity or calibrated market impact.

Diagnostic, incomplete evidence · 2025–2026 attribution; 2016–2026 benchmark

Style covariance explained part of a model return series. A separate equal-weight benchmark diagnosis exposed universe, new-listing and return-clipping defects.

Covariance is not causal skill attribution. Revised benchmarks still zero-filled missing stock-days and omitted costs; they cannot establish model promotion.

Historical exploration · 2015–July 2026; development through 2023

Raw, residualized and combined screens gave different results and large drawdowns. Public candidate and cadence tables are retained below.

The repeatedly inspected 2024+ segment is descriptive. Narrative and snapshot versions differ; do not splice them into a single final backtest.

Historical cross-section · 2008–July 2026; 223 formations

Point-in-time turnover and prior-window means or medians answer different questions. Quintile diagnostics are separate from later decile controls.

Alternative definitions require their own executable comparison. Definition sensitivity alone does not identify a unique premium.

Conditional association · Early 133 formations; later 86 common-support months

Later matched-stock controls retained a low-turnover association after size, risk, liquidity and return controls. A regression check also remained positive.

Earlier anatomy tables use a different sample and ranking basis. Neither double sorts nor regressions identify a causal or executable premium.

One-month observation · One completed month: August 2026

The frozen July-end signal produced opposite raw and controlled spreads in August. Only one completed month is reported by the reviewed source.

Do not mark later pending months complete merely because time has passed. Require actual receipts and multiple new observations.

Simulation; partial coverage · Separate daily portfolio and intraday samples

An explicit suspension correction lowered the daily simulation’s annual return from 7.04% to 5.33%. Minute coverage was about 80.8%; quote-level coverage was lower.

Missing quotes are not proven suspensions or zero returns. Incomplete minute and order-book records cannot support a full-history fill-cost replay.

Historical sensitivity · Public snapshot: 3 windows and 16 grid cells

The complete published grid is available, alongside three fixed monthly low-turnover windows. Definitions and engine scope differ between the tables.

A source caveat says the grid was pending while the same snapshot contains 16 rows. Preserve the inconsistency and inspect actual rows; do not infer independent confirmation.

Unreconciled comparison · Monthly and biweekly ledgers

Weight-level, ideal-share and continuous-ledger results differ. Formation counts, target counts and owner replay periods also differ.

Do not splice return series or declare one canonical result. Zero modeled impact is an assumption, not evidence of zero realized cost.

Path-dependent simulation · Seven public capital scenarios; separate lot checks

Cash, lot rounding and delayed orders changed outcomes. The published capital sweep has rising returns at larger capital.

This is not an economic capacity curve: deferred orders can skip later losers, and impact is uncalibrated. Opportunity-cost figures are diagnostic.

Source-reported historical results · 2015–July 2026; selected monthly 40-name results

Removing different bottom-cap fractions did not produce a monotonic improvement; large drawdowns remained.

A selected scorecard is not completion of the full weighting, buffer, cost and capacity protocol. Restricted configurations are not republished.

Historical bundles recovered; partial provenance · Different 2015+ and 2023+ samples

The dated monthly 100-name bundle was recovered through a documented storage migration; its stored full-sample summary agrees with the rounded source-note figures. The weekly 200-name proxy remains a separate comparison.

No performance was rerun. The recent-window comparison receipt and immutable code/input/output binding remain missing. A separate weekly proxy cumulative-excess label still fails percentage-point subtraction; do not pool or reinterpret these returns.

Outputs recovered; structural checks passed · Bottom 0/10/20/30%; weighting and buffer controls

The migrated bundle contains 112 factor rows, 56 weighting comparisons, 80 buffer rows and 24 capacity rows. The factor, buffer and capacity grids have unique keys, no blank cells and complete declared axes; four checkpoint files match final outputs byte-for-byte.

The recorded owner-code commit fields are unknown. This artifact audit does not rerun performance, authenticate every historical input, establish an untouched holdout or promote a strategy. The separate characteristic-decomposition protocol still has no located result bundle.

Protocol only · Size, illiquidity, extreme returns, volatility and quality

Double sorts and regression designs would test which characteristics accompany the apparent micro-cap effect.

No verified outputs were found for this protocol. Mechanism claims remain open, including delisting and shell-value effects.

Replay blocked · 2015–September 2026 validity audit

Earlier targets used future quote availability and current membership. Removing these filters changes the sample and invalidates direct reuse of old results.

A current membership snapshot cannot reconstruct historical membership. Missing held-security marks block a complete validated NAV.

Data repair, not return evidence · Historical-status supplement, September 2026

Historical names replace current-name ST inference. Strict date parsing exposed nullable export strings and previously admitted delisted securities.

A documented nullable adapter repairs representation only. Suspension-event matches do not prove a tradable price or settlement value.

Targets rebuilt; NAV unverified · Protocol frozen before new performance inspection

Historical eligibility is reconstructed without future-price filters; unknown status stays unknown. Old-target audits and new targets remain separate.

Targets are not a validated account. Do not drop missing-price names and renormalize; engine installation and run verification need separate receipts.

Incomplete valuation · Multiple audits with different windows

Public audits found 189 missing marks on 117 dates in one window; an older broader replay found 3,708 marks on 1,407 dates.

Counts are neither distinct delistings nor comparable rates without denominators. Remaining-name averaging silently renormalizes exposure; the return bias is not quantified.

Historical return and risk · 2008–2014 extension; later overlapping sources

The longer history adds earlier drawdowns and prolonged underwater intervals. Cleaned and earlier source series are shown separately.

Do not stitch sources without an overlap audit. An unfinished recovery is censored at the sample endpoint rather than assigned a completed waiting time.

Reference comparison · Vendor series, available membership and proxy indices

Published index points support return comparisons, while current membership snapshots lack historical effective dates and weights.

Dividend conventions, pool differences and incomplete daily data prevent a certified index replication. A proxy index is a separate benchmark.

Short-sample mechanism check · Public diagnostic dated 2026-09-28

Continuous holdings and cash expose effects of participation, T+1, price limits and assumed fees. The public bilingual source is included in the appendices.

This diagnostic does not certify full-history performance or actual capacity. A later method does not retroactively validate earlier ledgers.

Descriptive, quality-gated · A-share, Hong Kong, Japan and US; unmatched dates

Market-cap quantiles and USD ADV20 provide context, but dates, security types, units and eligibility differ. Some A-share totals do not reconcile.

Exclude inconsistent slices from rankings. Missing P90 cannot be replaced by a median; mean-versus-median lineage remains unresolved.

Partial coverage · Ten-date book audit plus separate historical trades

Published valid and excluded order-book counts reconcile, but missing minutes cluster on one date; a trades-only sample has no orders.

Book counts are not actual-order capacity calibration. Require aligned orders, trades, timestamps and observed fills before estimating implementation cost.

Promotion held · 2008–2026 matched raw, constrained and net comparisons

Data checks passed, but only half of the core factor gates passed. Constraints and costs reduced apparent factor returns.

Theoretical shorts, terminal proxies and reconstructed point-in-time fundamentals remain limits. This does not rerun or validate every earlier long-only strategy.

Infrastructure only · Adjacent factor-framework infrastructure

A small-cap sensitivity CSV is part of an output contract, with synthetic pipeline checks.

A file contract and synthetic test are not empirical small-cap evidence. No performance conclusion is added for this infrastructure.

Diagnostic, source-reported · Archived monthly comparisons, March 2026

Provider and point-in-time fundamental portfolios differed substantially. The provider result accompanied small-cap and short-horizon price exposures; size-neutral and softer controls changed that exposure.

Style correlation does not establish causality. Better-controlled size exposure did not establish stronger structural alpha, and this Hong Kong sample cannot validate A-share micro-caps.

Reconstructed proxy; substantial limits · September 2023–July 2026; 150 weekly formations

The strict historical-status window produced four selections at every formation. The standalone concentrated sleeve was highly sensitive to trading costs and experienced severe drawdowns; common-week index comparisons showed only partial agreement.

This is a reconstructed proxy, with revision safety false, current master-data inputs and incomplete limit-price execution modeling. Historical ST coverage ends in July; do not extend the strict sample through August or call it index replication.

Source-reported; share with caveats · Three-year studies; broad proxy and separate four-stock sleeve

The broad proxy comparison and later unified four-stock sleeve replay are distinct studies. A shared weekly calendar, next-session open execution and portfolio-level costs permitted a combined-account comparison.

The combined outcome does not isolate a micro-cap premium. Reconstructed fundamentals may contain later revisions; large drawdowns, concentration and incomplete execution constraints prevent live-capacity or formal point-in-time claims.

Cached exploratory evidence; not rerun · Cached notebook comparisons: 2025 and partial 2026

A concentrated candidate pool could not consistently satisfy portfolio size and beta constraints. Widening the candidate pool improved feasibility, while hard exposure controls reduced historical selected-sample excess returns.

The notebook reads existing comparison files and retains cached outputs; execution counters do not authenticate a fresh run. Samples were reused, lineage was incomplete and missing-outcome weights were flagged. Neither independent micro-cap alpha nor untouched out-of-sample performance is established.

Inspected; one family added

Seven source notebooks were found, all tracked in the two research repositories. Eight retained notebook files were also inspected: three are byte-identical copies, five are separate executed reviews. No additional untracked notebook was found in these two repositories.

Cached size/beta feasibility evidence is now H35. Empty launcher notebooks, duplicate copies and synthetic calibration outputs do not count as independent empirical confirmations. No notebook was executed by this audit.

53 rows matched; 60-session metrics replayed

All 53 published rows in six table families match retained CSV fields at numerical tolerance 1e-10. This covers candidates, anatomy, capital, cadence, ledgers and the joint grid. Separately, 2,778 retained suspension-corrected daily observations reproduce all four published 60-session-window metrics after rounding: 5.334535% annual return, 0.330517 Sharpe, −52.389378% drawdown and 6.785942 annual turnover.

The 20- and 120-session historical daily artifacts and a jointly bound three-window runtime remain missing. The 60-session metric replay establishes a retained daily source, not the original historical code, inputs or execution identity. A new current-inputs run matches the 20/60 rounded metrics but disagrees for 120 (4.22% versus 7.60% annual). It is a separate observation, not the missing historical binding; see the window recheck above.

Recovered through migration

The dated bundle and its self-referencing metadata were located after following a documented directory rename. All full-sample summary figures agree with the rounded source note; this closes the earlier blanket “bundle unavailable” gap.

The recent-window comparison receipt and immutable execution binding remain missing. Recovery of stored results does not reproduce performance, verify index membership or certify implementation.

Bundle recovered; provenance incomplete

The full named artifact set and declared factor, weighting, buffer and capacity grids were recovered. Unique-key and completeness checks pass for 112 factor, 80 buffer and 24 capacity rows. Four checkpoint CSVs match final files.

The owner-code commit fields are unknown. Input-vintage authentication, matching historical code and an independent rerun remain requirements before stronger reproducibility or strategy claims.

Hashes agree; valuation remains blocked

The stored readiness receipt and both target files match the hashes in the final verification receipt. Missing-mark table counts also agree with the stored audit.

The receipts explicitly say no NAV was published. Suspension matches do not supply settlement prices or prove fills, and this audit did not verify a new engine installation or replay.

Exact bundles still missing

The exact historical Top800 attribution JSON and yearly CSV, its run-linked factor matrix, and the characteristic-decomposition output set were not found in the inventoried warm roots, migrated research subtree and identified cold archives.

Similarly named later matrices cannot substitute. These are bounded negative findings, not proof of deletion. Resolution needs an exact artifact or a documented historical run/code/input binding; a new run must be labeled separately.

Executed reviews located

Later executed notebooks preserve descriptive liquidity alignment and conditional execution-cost reviews. Synthetic calibration notebooks are separately identified as infrastructure evidence. The inspected cold-storage volume was available.

Later aligned samples do not erase older source discrepancies. Numeric notebook outputs with unresolved redistribution permission remain local; cached execution and synthetic tests do not establish actual-order capacity.