Deep-learning market analysis for a fintech

Challenge

The client's analysts were building signals by hand in spreadsheets, with no reproducible backtesting and no way to combine price data with news sentiment.

What we built

• A data pipeline pulling price, volume and news feeds into a versioned feature store.
• Sequence models (LSTM/temporal CNN) for short-horizon trend classification, combined with a transformer-based sentiment scorer over news headlines.
• Walk-forward validation with transaction costs, benchmarked against a simple momentum baseline.
• Dashboards showing signal confidence, exposure and drift, and a written model card describing limits and failure modes.

Results

In walk-forward testing the model-guided allocation improved returns by about 22% relative to the client's baseline strategy while reducing drawdown. More importantly, the team now has a repeatable research process instead of ad-hoc spreadsheets.

Financial models are probabilistic tools, not oracles. Every report ships with its confidence intervals.