Danelfin publishes an audit of its own AI Score. Here is what that kind of track record can and can't prove

Danelfin rates stocks from 1 to 10 every trading day with what it calls an AI Score. Its own footer describes the score as a calculation of 'stocks' probabilities of beating the market based on past market behavior', adding that it 'does not calculate or predict actual results' and that illustrative portfolios are backtested. The company also runs a separate site, Danelfin AI Audit, that tests the score with dozens of statistical checks. We read the method, not the marketing numbers, on October 14, 2026.

What the score is

According to the audit's own description, each trading day every stock receives a score based on about 900 technical, fundamental and sentiment indicators. The universe can be set to all US stocks (about 5,600), the S&P 500 or the Russell 2000. The model optimizes for a 3-month forward window. Plan details are on our Danelfin tool page.

The one sentence every reader should find

Repeated across the audit's tabs is this line: scores before July 7, 2025 'were recalculated walk-forward by the current model (simulated backtest); scores from Jul 7, 2025 onward are live, point-in-time.' In plain terms, roughly eight years of the history the dashboard covers (it splits results into 2017 to 2019, 2020 to 2022 and 2023 to 2025) were produced by today's model looking back, and about fifteen months are live records of what the score said at the time.

That distinction is the whole game for any AI score. A recalculated history can be built carefully, without peeking at future prices, and still benefit from a model designed with knowledge of what worked. Only the live period is evidence of the kind a buyer should weigh most.

What the audit does well

Unusually for a vendor, the dashboard runs checks designed to embarrass a weak signal, and labels their limits:

• Entry the next day. Returns are measured from the day after the score date, to avoid same-day look-ahead.
• Multiple-testing hurdle. It tests whether t-statistics clear the Harvey-Liu-Zhu threshold of 3.0, meant to account for data-snooping across hundreds of published anomalies.
• Factor adjustment. It compares raw alpha with alpha left after Fama-French factors and momentum.
• Costs, capacity and borrow. It models trading costs, market impact at larger portfolio sizes, and the borrow cost a long-short version would pay.
• Survivorship and delistings. It tests sensitivity to how delisted stocks are treated.

What the audit itself says it cannot do

The most useful text on the site is its own caveats:

• The sub-period comparison 'is a retrospective sub-period consistency check, not a true out-of-sample test.'
• Its out-of-sample R-squared is out of sample 'for the regression only', not for the scores, and the date split 'does not purge overlapping forward-return horizons', so 'this is not a guaranteed prospective evaluation.'
• The bias checks 'cannot detect look-ahead inside the scores themselves', and the score-file consistency test 'cannot detect restatement by Danelfin.'
• The long-short strategy 'assumes zero borrowing cost for the short leg' unless you adjust it.
• A dedicated note explains that a score of 10 describes a group's average tendency and 'does not guarantee this particular stock will' beat the market.

How to read any AI stock score's track record

Danelfin's audit gives a checklist that works for every service selling a score:

• Find the live start date. Separate recalculated history from scores recorded at the time. Weigh the live period most, and note how short it may be.
• Ask whether the model changed. If a current model rescored the past, the past was scored with hindsight about which features mattered.
• Check the unit of the claim. A score describes buckets of many stocks over many dates, not the next move of the one stock you are looking at.
• Net out costs. Turnover, spreads and borrow fees come before any edge reaches you.
• Look for a monotonic pattern. Higher scores should line up with better outcomes across all ten buckets, not just at the extremes.

What we leave out

The dashboard displays equity curves, alpha by score and risk-adjusted ratios for its model portfolios. We do not reproduce them: most of the period is simulated, and our rule is not to showcase backtest gains. Readers can explore the live-only window on the audit site by setting the date filter to start July 7, 2025.