RavenPack Introduces NVIDIA Nemotron-Powered Pipeline for Backtest-Ready Investment Datasets on Bigdata.com

RavenPack fine-tuned a 4B NVIDIA Nemotron model to label more than 845,000 financial-text excerpts in under an hour on two NVIDIA H200 GPUs, enabling full-corpus dataset generation for financial research.

BERLIN, Oct. 21, 2026 /PRNewswire/ -- RavenPack, the data and AI infrastructure provider to the world's leading hedge funds, banks and asset managers, today introduced a new dataset generation pipeline on Bigdata.com, its AI grounding and search platform. The pipeline turns a research hypothesis expressed in plain language into a labeled, backtest-ready dataset built from the full retrieved corpus. RavenPack's pipeline, powered by a compact NVIDIA Nemotron model, labels financial text at production scale.

Why it matters now

Every data science team has more ideas than it can test. Advanced AI models can interpret financial text in ways that support analyst-like research, but running them across millions of documents and a decade of history can be slow and expensive. As a result, most ideas get tested on a sample, or not at all. At the same time, sentiment has become a widely used input, and funds are looking for more specific signals that tell them why a company's story is changing.

RavenPack's pipeline addresses both challenges. It makes large-scale labeling more practical to repeat, and produces signals that go beyond sentiment. In RavenPack's initial backtest, the pipeline outperformed a relevance-weighted sentiment-only strategy built using the same portfolio construction framework.

From hypothesis to signal: a worked example

The first test asked a simple question: is a company gaining or losing market share against its competitors?

The pipeline retrieved 845,220 excerpts from news and earnings-call transcripts covering 1,642 companies in a point-in-time Top US 1000 universe by market capitalization. It labeled every excerpt against a purpose-built taxonomy, then converted the labels into a daily, company-level signal. That signal was backtested from January 2016 to October 2026.

Two steps make this possible at scale.

1. Targeted retrieval. The Bigdata.com search API, with bigdata-smart-batching, retrieves relevant evidence from a licensed archive of more than one billion documents spanning 20+ years of news and earnings-call transcripts. Rather than issuing a query for every company in every period, smart batching plans retrieval around content volume and co-mentions, cutting the query count by more than 90% compared with a full company-by-period grid.

2. Efficient labeling through distillation. Larger reasoning models can classify financial signals, but applying them across an entire retrieved universe can be slow and costly. RavenPack used NVIDIA Nemotron 3 Super 120B to label a stratified sample of 84,521 excerpts, then used those labels to fine-tune NVIDIA Nemotron 3 Nano 4B using LoRA supervised fine-tuning. Running on two NVIDIA H200 NVL GPUs, the distilled model labeled the full corpus at both chunk and sentence level. In RavenPack testing, it achieved approximately 39x the throughput of the teacher model.

By the numbers in RavenPack testing and analysis:

• Speed: 845,220 passages labeled in 0.7 hours (about 49 minutes), roughly 39x the throughput of the 120B teacher model.
• Cost: Labeling the full history costs 63% less than a proprietary model at the same performance level, including the one-time cost of distillation. Each subsequent pass costs 96% less.
• Fidelity: On data held out from the teacher labels, the 4B student matched the teacher on 94.05% of chunk-level labels and 95.90% of sentence-level labels, with macro-F1 scores of 0.831 and 0.866 respectively.
• Signal quality: With smoothing over a 1-day half-life, the sentence-level signal produced an information ratio of 1.18 and annualized excess return of 8.63%. The chunk-level signal produced 0.91 and 6.58%. The same portfolio construction using relevance-weighted sentiment alone, with no classification labels, produced 0.41 and 3.78%.

Why open weights mattered

Distillation means training a smaller model on a larger model's outputs, then fine-tuning the results for a specific task. That requires direct access to model weights, so the model can be fine-tuned and deployed in an organization's own environment. Nemotron's open weights let RavenPack fine-tune once and run repeated inference workloads from then on, rather than relying on per-token access. For regulated institutions, the model can be run in-house, so its labeling logic can be inspected and audited.

"Data science teams are not short of ideas. The bottleneck has always been turning those ideas into reliable datasets with sufficient coverage and history. This pipeline changes that: clients can move from a hypothesis to testing it across the full breadth of content indexed by Bigdata.com," said Peter Hafez, Chief Data Scientist at RavenPack.

Availability

The pipeline is available today to Bigdata.com clients. Developers can run it end to end, from idea to labeled dataset to signal and backtest, in the companion notebook, From a research idea to a backtested signal.

About RavenPack

RavenPack is a leading provider of data and AI infrastructure for financial institutions, with a 20+ year track record serving the world's most sophisticated hedge funds, banks and asset managers. RavenPack operates two core products: RavenPack EDGE, its alpha and risk analytics platform, and Bigdata.com, its AI grounding and search platform.

About Bigdata.com

Bigdata.com, a RavenPack product, is the grounding layer for AI: a precision search and retrieval platform and token-denominated clearing house that lets AI agents pay only for the exact content they consume, cutting token consumption by orders of magnitude while grounding every answer in licensed, cited sources. Natively connected to leading AI assistants and accessible by any agent, Bigdata.com powers grounded, traceable answers for the agent economy.