olivlaw

Method

Research ledger: what we tested and what failed

Most ideas we test do not hold. We write them down the same way as the ones that do, because a hidden negative result makes the next positive one worth nothing.

10 tests listed here; 7 were not confirmed.

How a test is run

  1. The decision rule is written and sealed before the run: what is measured, on which period, what threshold means «confirmed».
  2. The verification period is looked at once. What has already been looked at is recorded, so it is not reused as fresh evidence.
  3. Trading costs are inside the figure. An effect that disappears after costs is a negative result.
  4. The verdict uses fixed words: not confirmed, undecided, correlated. «Causes» is never written.
  • BE-1not confirmed

    Can a ternary-weight embedding model (BitNet-embedding-0.6B) replace the production model on Romanian retrieval?

    Lower nDCG@10 on all three Romanian sets (−0.106 / −0.057 / −0.291, every interval below the −0.02 margin) at 0.25–0.46× the speed on 3 threads. Nothing was changed.

  • Q-11not confirmed

    Does buying only around the new moon add return or cut the losses of a stock portfolio?

    −1.44 percentage points a year against the same strategy without the rule (interval −3.09…−0.51); the worst month is identical.

    Replicates: Dichev & Janes 2003; Yuan, Zheng & Zhu 2006

  • SL-1open

    Do stocks with a large earnings surprise beat the index over 60 sessions more often than not?

    46 % (150 of 324, interval 41–52 %): mean +0.9 points, median −2.0. Below half, so we say so; it is now judged prospectively.

  • EL-1nothing certain

    Does the US stock market react differently depending on which party wins the election?

    S&P, 49 elections from 1928 to 2024: the Republican − Democrat gap on the next day is +0.51 points, t 0.91 — nothing certain. Firms follow the party they fund (t 3.21), but the result is fragile and is gone overnight.

  • L-4holds

    Does post-earnings-announcement drift still exist: the top decile of surprise against consensus, held for 60 sessions?

    On the validation period, +2.33 % against the universe (t 3.0). The only established strategy we tested that still holds.

    Replicates: Bernard & Thomas 1989

  • L-1…3not confirmed

    Do three effects from the literature still work: the volume premium, residual reversal and residual momentum?

    −0.24 % / −1.29 % (t −3.8) / −0.05 % a month against the universe. None is confirmed; the effects have decayed.

    Replicates: Gervais, Kaniel & Mingelgrin 2001; Jegadeesh 1990; Blitz, Huij & Martens 2011

  • E-3not confirmed

    Does the reversal after a volume spike, described on a small group of stocks, hold across the whole market?

    It reproduces on the original group (+8.7 %). Across the market it loses: −10.5 % a year for the reversal, −23.3 % a year for the continuation.

  • N-3not confirmed

    Do volume patterns alone (accumulation, breakout, a high on volume) predict returns?

    All negative after costs; the volume breakout loses 233 basis points in 20 sessions (t −4.5).

  • N-2not confirmed

    Do press stories and rumours come before large price moves?

    Only 6.7 % of moves of at least 8 % have a headline in the 10 days before; the anticipation lift is about 1, which is none.

  • EQ-P3not confirmed

    Does a stock that is «cheap» by our own valuation formulas beat the market?

    The rank correlation on the held-out period is negative (−0.09…−0.10). The valuation bands are not published as a signal.

The list is a selection: tests that revisit a published idea on public market data, plus model evaluations. It is not investment advice and no strategy here trades money.

The published forecasts have their own ledger, with a Brier score →

How to cite this page

Plain

olivLaw, «Research ledger: what we tested and what failed», November 19, 2026, https://olivlaw.com/en/us/learning/registru (accessed November 20, 2026).

APA

olivLaw. (2026, November 19). Research ledger: what we tested and what failed. https://olivlaw.com/en/us/learning/registru

BibTeX

@misc{olivlaw-2026-registru,
  author       = {{olivLaw}},
  title        = {Research ledger: what we tested and what failed},
  year         = {2026},
  howpublished = {\url{https://olivlaw.com/en/us/learning/registru}},
  note         = {Accessed 2026-11-20},
}

The texts are produced by olivLaw's AI pipeline; the author of record is the organisation.