DIVINE reconstructs 77 targets from 16 technical indicators for cross-market stock pretraining, achieving the strongest average portfolio performance across six markets with a 0.05M-parameter encoder

It proposes DIVINE (DIVerse INdicator rEconstruction), a cross-market pretraining framework that reconstructs technical indicators from raw OHLCV history for downstream transfer to stock ranking. Whereas existing financial time-series pretraining objectives learn from masked observations, contrastive relations, or future outcomes, this work uses technical indicators computed from observed price-volume history as supervision, aiming to avoid future-supervision uncertainty while maintaining alignment with return prediction. The abstract states that technical indicators are computed from observed price-volume history, are consistently defined across markets, summarize diverse market dynamics, and have established relevance to return prediction; pretraining is performed jointly on six equity market datasets.

Pretraining reconstructs 77 targets derived from 16 standard indicators, and only the learned encoder is transferred to downstream stock ranking. It operationalizes supervision design as a multi-indicator, multi-target reconstruction task and decouples pretraining from the downstream task by transferring only the encoder. The abstract gives the concrete counts of 77 targets and 16 standard indicators and states that only the learned encoder is transferred to downstream stock ranking.

Across all six markets, DIVINE achieves the strongest average portfolio performance with a 0.05M-parameter encoder, outperforming pretraining baselines and matching or exceeding substantially larger financial foundation models while remaining robust and data-efficient. A lightweight encoder reaches this comparative result on cross-market average portfolio performance, pointing to supervision design and cross-market diversity rather than model scale as key drivers of strong transferable financial representations. The abstract reports an average portfolio performance comparison across six markets involving pretraining baselines and substantially larger financial foundation models, and describes the results as robust and data-efficient; specific numbers are not given in the text.

Systematic analyses show that indicator diversity and market diversity provide complementary gains in transfer. It decomposes the source of performance into supervision-target diversity and cross-market data diversity, indicating that the two are complementary. The abstract presents this as a finding of 'Systematic analyses' without listing specific ablation settings or values in the text.

The work targets cross-market stock ranking: joint pretraining on six equity market datasets, transfer of only the encoder to the downstream ranking task, and average portfolio performance as the evaluation lens. Its scope is markets with raw OHLCV history and consistently definable technical indicators; the abstract's claims of robustness and data efficiency are likewise bounded by this pretraining-transfer setting. For researchers and practitioners seeking transferable financial representations from a lightweight encoder, the framework offers a path that substitutes supervision design for model scale.

The text is abstract-level information: it does not give per-market portfolio performance values, the time periods of pretraining and downstream data, the baseline list and their scales, or the quantitative basis for 'robust and data-efficient'; the specific ablation settings behind the complementary gains of indicator and market diversity are also not listed. Readers should therefore still watch whether these conclusions hold across different markets and evaluation lenses, and under what conditions the comparison between the 0.05M-parameter encoder and substantially larger financial foundation models holds.