Entering at the open leaves almost none of the commercial models' news signal to trade. The three commercial news readers have daily ICs of 0.074 to 0.077 against the overnight gap. Against the open-to-close return, the range is -0.001 to 0.002.
Only the latter pays an investor entering at the open.
The morning-news bet
The study starts with a familiar premise: prices underreact to company news, leaving a move to earn after the open. A model reading headlines before 09:30 should be able to anticipate it. Dehkordi, Kargarzadeh, the two Khaledians, Parvini and Ghatak test that proposition on point-in-time S&P 500 members.
A learned importance filter selects company-sessions with relevant news. Each gets a JSON state assembled as of 09:15 ET, stripped of names, tickers, people, dates and every percent price move. The decision model returns probabilities over typed answers rather than generating text. It scores five-day relative performance across nine levels, and the probability-weighted level supplies the news-implied move. Readings at 10:00 and after the close then judge whether the price reaction has been too small or too large.
Laya-Fin, the primary model, is a ModernBERT-large encoder fine-tuned on July 2024 to June 2025 with one seed and one epoch. It is tested alongside three commercial models (d1-ultra, span-02, drex-v2.0) and a set of baselines. OLIVIA, the authors' protocol, confines selection to two 2025 validation quarters. The authors freeze the analysis code with a SHA-256 hash, then unblind the January to July 2026 test once. It contains 13,748 pre-open events across 140 sessions. They separate overnight from intraday abnormal returns at the open, remove market and sector moves, and scale each leg by a 126-session median-absolute-deviation volatility.
For Laya-Fin, none of the five primary hypotheses survives Holm correction: adjusted p ranges from 0.269 to 0.853. No pre-registered book clears both costs and the deflated Sharpe ratio.
Which return was still available?
The precise split at the open deserves copying. The overnight leg runs from the previous close to the official open, though some of that return has already printed when the model reads the news. Pre-market trading begins at 04:00; the reading comes at 09:15. The authors acknowledge this and leave the overnight leg unsplit at 09:15.
Their timing cut makes the problem visible. For d1-ultra, overnight IC reaches 0.115 (t = 5.0) when the latest article arrives between 08:00 and 09:15. It is 0.061 for articles arriving from 04:00 to 08:00, and an insignificant 0.024 before 04:00. Reading news while the market is already moving makes some of the measured relationship co-movement with price discovery. An implementer with pre-market quotes should measure the 09:15-to-open leg separately. The tradable side here is empty.
Validation gave way on the test
Laya-Fin's close-to-close IC fell to 0.020 (t = 1.67) on the test. It had been 0.055 in the first validation quarter and 0.065 (t = 3.64) in the second. The test result falls below the validation interval's 0.030 lower bound.
The post-open signal deteriorated most. An IC of 0.054 (t = 2.89) in the final validation quarter became 0.010 (t = 0.82) on the test. The commercial models also lost roughly half their IC, with d1-ultra falling from 0.097 to 0.039.
The authors' exploratory post hoc split locates the surviving signal in scheduled earnings reports. Across 1,128 such test events, Laya-Fin had an IC of 0.134, following 0.125 and 0.106 in the two validation quarters. Its IC on all other test events was 0.013. Earnings days are marked as seen in the paper: the authors' AI research assistant had examined their test values before the exploratory study was designed. The aggregate IC therefore combined a stable earnings effect with a fading remainder. A replication should separate them from day one. Even on earnings days, the signal arrives before the open: d1-ultra's test IC is 0.238 overnight and 0.034 afterward.
The conditioning study gives the strongest warning. Just 8 of 75 model-condition pairs met the requirement for a positive t of at least 1 in both quarters and a pooled t of at least 2. On the test, 5 changed sign. Tradable versions had reached in-sample gross Sharpe ratios of 6.8 on val-B; all lost after 10bp.
Costs finish the main books
The one-day books turn over about 4.4 times capital daily. Laya-Fin's open-to-close book earns an annualized gross Sharpe of 0.35, with a break-even cost of 0.5bp per side. Charging 10bp per side on both legs and the SPY hedge takes it to -6.03 net.
The highest gross Sharpe comes from trading only Laya-Fin's most confident pre-open calls and holding until the next close. That book earns 1.88 gross and -0.38 net. Its break-even cost is 8.3bp, and it reaches the 93rd percentile of the random-signal placebo.
At 10bp a side, these books lose even before search is accounted for. The deflated Sharpe ratio imposes a tougher hurdle. Across 545 logged trials, the convergence strategy faces a best-of-null benchmark of 0.26 per day. No book has a deflated-Sharpe probability above 0.024. The book achieving that 0.024 holds one name a side on a handful of days, with net 1.50 and turnover 0.05. None of the net figures includes borrow, so all are flattering. The model-based convergence book also holds the same positions as a price-only gap fade 92% of the time, with a daily correlation of 0.96.
Company masking leaves room for recognition
Contamination remains difficult to dismiss. The screened commercial models could identify masked companies from four same-sector candidates 62% to 74% of the time, compared with a 28% matched baseline. d1-ultra scored 74%. All were released in October 2026, after the sample.
Two findings temper that concern. The untuned Laya model identified only 32% of companies and still produced the overnight split (H1 = 0.054, p = 0.001). Restoring company identities for d1-ultra on 5,923 events changed its IC only from 0.026 to 0.029. The authors further disclose that five of them wrote an earlier study with a sample overlapping early 2026. A test spanning 140 sessions leaves small ICs difficult to distinguish from zero in either direction.
The cleanest check on memorization is the forward test. Pre-registered on 29 October 2026 for news after 21 October, it has yet to run.
Why we could not rerun the 2026 test
We lack the private NewsWitch corpus and its per-article availability timestamps. We also lack the exact model versions, historical index membership and the 2026 window. Running a frozen classifier on platform news for a top-500 capitalization screen would test a related idea with different inputs. It would tell us nothing about these models.
Until the forward window reports, take the split at the open into your own backtests. It costs nothing to calculate, and it determined this paper's headline result.