We published the paper. Now we're building it.
QuantPrompt started as a research question: what happens when you let a language model read the output of a time-series forecast, alongside market sentiment and technical indicators, and ask it to make the call and explain itself? The answer is now out in Springer Nature. We're turning it into something people can actually use.
A forecast you can argue with.
Most prediction models hand you a number and leave you to trust it. We wanted the opposite: a system that takes the forecast, weighs it against what the market is saying and what the charts show, and then writes down its reasoning. You get a direction and the argument behind it. If you disagree with the argument, you can see exactly where.
How it works
Forecast
Recurrent neural networks (LSTM and GRU) are trained on price history to forecast the 30-day trend, then reduced to a simple up-or-down signal.
Add context
Reddit sentiment, scored with BERT, and classic technical indicators (RSI, MACD) are lined up next to the forecast so it isn't judged in a vacuum.
Reason
A language model reads all of it and produces a final directional call with a written rationale. Every call is logged, so you can go back and check its work.
Where the signals live.
The Hub is the place we're building to hold everything the model produces: each signal, the inputs that fed it, and the reasoning it gave, all timestamped. Early days, but it's the part we're most excited about. More on that when it's ready to show.
The paper
Peer reviewed at the 33rd Irish Conference on Artificial Intelligence and Cognitive Science (AICS 2025, Dublin), published by Springer Nature in Communications in Computer and Information Science, vol. 2950.
Bitcoin Prediction: A Hybrid Approach with Recurrent Neural Networks and GPT-Based Reasoning
Zhengri Cui, Daniel Comerford, Martin Crane, An Nguyen, Marija Bezbradica
School of Computing & ADAPT Centre, Dublin City University
78.9% F1 on direction
The hybrid model called the 30-day direction more accurately than any single model we tested, including the forecasting networks on their own.
194% over buy-and-hold
In a trading simulation over May 2024 to May 2025. One test period, so read it as promising rather than proven.
It shows its work
The reasoning layer outperformed the raw forecasts and, unlike them, left a written rationale behind for every call.
Where it's at.
The signal engine from the paper is built and running. What we're working on now is the next layer: a set of agents that aggregate the individual signals, and a dashboard that presents the result in a way an asset manager could actually sit in front of. Nights and weekends, no rush, no funding round. Just building.
Say hello
Curious about the research, want to try it, or think you could help? Drop us a line. We read every one.
Thanks. Got it.
We'll reply at the email you gave us.