←BaldanAll worksLet's talk

Case study · Kalla Group · 2023

Forecasting bitumen prices with a neural network.

Bitumen is bought in bulk and its price moves with oil. This began as my thesis and became a dashboard stakeholders check: data from several subsidiaries brought together, a recurrent neural network forecasting the FOB Bitumen Singapore price, and the result on a page anyone can open.

Big Data AnalystRNN · time seriesStreamlit · Plotly
FOB Bitumen Singaporeillustration
historyforecast

The notebook and data aren't part of this portfolio, so this page sticks to the approach.

01 · From data to a decision

Four steps from scattered data to a forecast.

  1. 1

    Bring the data together

    Records from several subsidiaries, merged with SQL into one clean series.

  2. 2

    Shape the series

    Prices ordered in time and cut into windows: the past values the model sees, and the value it has to predict.

  3. 3

    Train the network

    A recurrent neural network learns from those windows, carrying what it saw earlier forward through the sequence.

  4. 4

    Put it in front of people

    A Streamlit app with Plotly charts shows the history and the forecast, so the forecast is used, not filed.

02 · Why a recurrent network

A model that reads prices in order.

  • Order matters

    A price series isn't a bag of numbers: yesterday says more than last year. A recurrent network reads the window step by step and keeps a running memory.

  • Windows, not single points

    Each training example is a stretch of past prices and the price that followed, so the model learns patterns over time.

  • Checked on unseen time

    A forecast model is judged on a later period it never trained on, the way it will be used.

  • A dashboard, not a report

    Stakeholders open the app and see the latest forecast next to the history, instead of waiting for a document.

03 · The small stuff

What it took beyond the model.

Data from many places

Several subsidiaries, several formats, one series.

SQL first

The merge and cleaning happened in SQL before any modelling.

Interactive charts

Plotly charts people can zoom and hover.

From thesis to tool

The research became something people use.

04 · How it fits together

From databases to a dashboard.

Subsidiaries

Source data

Records from several companies in the group.

Analysis

SQL + Python

Merging, cleaning, windowing and training the recurrent network.

Web

Streamlit app

History and forecast in Plotly charts for stakeholders.

05 · Tech stack

What each piece does.

SQL

Merging the subsidiaries' data.

Python

Preparing the series and training.

Recurrent neural network

The forecasting model.

Streamlit

The dashboard app.

Plotly

Interactive charts.