Building practical deep learning skills since 2020

We developed an online platform focused on helping professionals apply neural networks to real price forecasting problems through structured practice and experimentation.

Our Foundation

Started with a specific technical gap

In 2020, we noticed something while working with teams trying to implement forecasting models. Most training programs covered theory well enough, but participants struggled when applying LSTM networks or transformer architectures to actual price data. The gap between conceptual understanding and functional implementation was wider than expected.

We built Temqora to address this particular problem. The platform focuses on structured practice with real datasets, step-by-step model construction, and debugging common issues that appear when training on financial time series. Each workshop walks through the full pipeline from preprocessing price sequences to validating predictions against holdout periods.

Our approach came from direct project experience. We found that learners needed hands-on sessions with immediate feedback loops, not just video lectures. The workshops include exercises where you implement backpropagation for specific architectures, tune hyperparameters based on validation metrics, and handle problems like gradient vanishing or overfitting on trending data.

The learning environment replicates actual development workflows with Jupyter notebooks, version control for model experiments, and collaborative debugging sessions.

Since launching, we've refined the workshop structure based on where participants get stuck. Common issues include handling non-stationary price series, selecting appropriate loss functions for regression tasks, and interpreting attention weights in transformer models. Each module now includes specific exercises targeting these pain points with guided solutions.

The platform serves professionals across South Africa's different regions through remote access, allowing learners to work through assignments at their own pace while maintaining scheduled sessions for collaborative problem-solving. Progress tracking shows completion rates and identifies concepts that need additional practice time.

Workshop session showing neural network architecture implementation

Collaborative debugging session for model optimization

2020

Platform established with initial workshop curriculum

100+

Practical exercises covering model implementation

How the learning structure works

Each workshop follows a progression from basic model setup to production-ready forecasting systems, with exercises designed around actual implementation challenges.

Code-first implementation

Build models by writing PyTorch or TensorFlow code directly. Exercises start with simple feedforward networks and progress to recurrent architectures and attention mechanisms, with each step producing functional prediction models.

Real financial datasets

Work with actual price series including stocks, commodities, and currency pairs. Assignments cover data cleaning, feature engineering from OHLCV data, and handling issues like missing values or outliers in historical records.

Debugging and optimization

Practice identifying why models fail to converge, why predictions lag actual values, or why validation loss diverges from training loss. Sessions include profiling training speed and reducing inference latency for production use.

Collaborative problem-solving

Scheduled sessions where participants share model architectures, compare training strategies, and work through technical roadblocks together. Discussion focuses on practical trade-offs rather than theoretical comparisons.

Incremental skill building

Modules progress from univariate prediction to multivariate models with exogenous features. Later sections cover ensemble methods, uncertainty quantification, and backtesting strategies against historical performance.

Validation and metrics

Learn to evaluate forecasts using appropriate error metrics for financial data. Exercises include walk-forward validation, calculating directional accuracy, and assessing prediction intervals rather than point estimates.

View workshop curriculum