Not supported · Outside indicators & papers · 2026-10-01
Does the 'Hybrid AI trading system' paper (SSRN 6139568) reproduce?
What this means
Built honestly, the model predicted next-day direction 51.3-51.6% of the time, no better than always guessing 'up' (52% of days were up days). A one-day labelling slip, using that day's own move as the label, lifts accuracy to about 74%, so a leak of that kind is the likely source of the paper's 63%. Even with hindsight and no costs the system made +71% in 2023-24 against +87% for holding the same stocks, and it trailed buy & hold in every universe and period tested.
The explanation is our best reading; the data shows the result but not the cause directly.
By timeframe
| Timeframe | Verdict | n | Result |
|---|---|---|---|
| 15-minute | Not tested: outside paper replication on daily data by design | ||
| 1-hour | Not tested: outside paper replication on daily data by design | ||
| 4-hour | Not tested: outside paper replication on daily data by design | ||
| Daily | Not tested: outside paper replication on daily data by design | ||
- Hypothesis
- XGBoost + technical score + regime gates gives 63% next-day accuracy and +135% over 2023-24.
- Method
- Algorithm 1 implemented in Python (XGBoost, features, entry/exit, sizing; next-open fills), adjusted Yahoo daily; three universes; honest (train 2019-22) vs paper-like shuffled split; 2023-24 and 2025-01 to 2026-09; 0 and 5 bp costs. FinBERT sentiment exit not tested.
- Result
- Next-day accuracy 51.3-51.6% (honest) vs 52% up days; an off-by-one label leak gives 74%. Best case (hindsight universe, leaky model, no costs) +71% in 2023-24 vs +86.7% for holding the same stocks; honest model 2025-26 +14.1% / +6.8% after costs vs B&H +47.7%. Trails buy & hold in every universe and period.
- How much to trust it
- Medium: faithful implementation, but sentiment exit omitted
- Did it change the live rule?
- not adopted
Scripts and data for this study will be linked here when the research code is published on GitHub.
Education and research only. Not financial advice. Backtests use today's index members (survivorship bias).