DG26
EN/PTContact
← Work02 / RESEARCH INTERFACE / FORECASTING2024—2025

FINANCE · MACHINE LEARNING / 2024—2025

B3Forecast

A forecast is useful only when its context stays visible.

B3Forecast is a research interface for exploring Brazilian market history and turning LSTM experiments into an inspectable forecasting workflow.

Personal project · research interface

60%less analysis time50+ active users · <100 ms ingestion latency

Reported project figures; evaluation limits are explained with the results below.

View live projectOpen source
My contribution
Personal research project connecting market history, LSTM experiments, and an interface for reviewing forecasts.
Key constraint
Keep asset, period, history, and model output visible without presenting a forecast as certainty.
What this case shows
A public project and source repository are linked below. The visual in this case is conceptual and contains no live data.

01 / CONTEXT

What needed to become clear.

A forecast presented as a single number can appear more certain than it is. The challenge was to keep the asset, time range, historical data, model output, and visual context in one readable path.

APPROACH

The interface puts context before prediction. Users select the asset and range, inspect historical data, and follow how the model turns that information into an interpretable output.

DOMAIN

RESEARCH INTERFACE / FORECASTING

02 / DECISIONS

The product logic behind the interface.

  1. 01

    Compact configuration keeps the main canvas dedicated to analysis.

  2. 02

    Historical data and model output remain in the same flow instead of hiding the origin of the forecast.

  3. 03

    Tables and charts are organized by analytical role, not only by visual type.

  4. 04

    A research interface instead of a simulated trading terminal.

  5. 05

    Reported latency below 100 ms in the documented scenario.

03 / SYSTEM PATH

A visible path through the system.

01Market history

Select an asset and time range.

02Model context

Follow the data into an LSTM experiment.

03Research interface

Keep history and model output together.

04Forecast review

Inspect the output in its original context.

04 / OUTCOME

B3Forecast reached a reported 87% prediction accuracy, processed more than 10,000 data points per day, and reduced analysis time by 60% versus manual methods. The project also supported 50+ active users and made the path from data to model to forecast observable.

Scope of the reported resultsThese are reported project results. This portfolio does not document the evaluation protocol, accuracy formula, or active-user time window; the figures should not be read as independently validated benchmarks.

What I learnedForecast interfaces need to communicate uncertainty and provenance, not only output.

STACK / TOOLS IN CONTEXT

PythonTensorFlowPandasLSTMStreamlit

CONTACT / THE NEXT SIGNAL

Have a complex product?

Talk to me about product architecture, data-rich interfaces, and creative implementation where the experience has to remain clear.