Skip to content

Boast FY23 SR&ED

Source: Notion | Last edited: 2023-11-22 | ID: 74448f17-5df...


  • Improve model performance
    • Experimented with using a new model structure (momentum-based transformer), and using not only dynamic features which change with time, such as price, volume, volatility etc, but also predictable/static features such as day of week, month of year, symbol category, etc enrich the model’s contextual understanding and performance.
    • Transformed the approach to loss functions in machine learning by implementing risk-adjusted returns instead of traditional absolute return metrics. This strategic shift aims to minimize potential losses while optimizing for realistic market conditions, where exposure control is crucial during major market drawdowns.
  • Development of Training Data for Second-Layer Models ◦ Orchestrated a sophisticated process for generating accurate training data for the EonPredictor, a second-layer model. Utilized a ‘rolling backfill’ method to circumvent the limitations of backtesting in machine learning models. This involved sequentially training models with historical data up to specific years and then applying these models to generate profit and loss (PNL) data for subsequent years, ensuring a realistic and incremental training approach. Chen’s work on improving model performance through the integration of a momentum-based transformer and the combination of dynamic and static features directly contributes to technological advancements in predictive modeling. This aligns with SR&ED’s focus on innovation and resolving technological uncertainties. The adaptation of loss functions to include risk-adjusted returns, instead of traditional absolute return metrics, represents a strategic shift in machine learning for financial applications, aiming to minimize losses while optimizing for real market conditions.
  • Dynamic Leveraging: an effective way on reducing maximum drawdown and increasing gains
    • It utilizes the Kalman filter running in real-time to estimate local mean in conjunction with a pre-trained logit curve to determine the optimal leverage.
  • Dynamic Weighting** :** dynamically weight the trading intervals (ie, 15m, 30m, 45m, 1hr, 2hr) using scipy.optimize based on the last N timestamps of data.
  • Evaluate Feature Split: evaluate the efficacy of a timeseries feature for Enigma performance enhancement. It does this by spliting the profit and loss returns into two categories and check if their distributions differ. The more they are different, the better efficacy of the feature

Software development services - hourly - 2022-11-28 through 2022-12-20 26.0 $2,600

File attachment

Software development services - hourly - 2022-01-30 through 2023-02-27 35.0 $3,500

File attachment

Software development services - hourly - 2022-02-28 through 2023-03-25 25.0 $2,500

File attachment

Software development services - hourly - 2022-03-27 through 2023-04-26 8.0 $800

File attachment

Software development services - hourly - 2022-04-27 through 2023-05-25 24.5 $2,450

File attachment

Software development services - hourly - 2023-05-26 through 2023-06-25 21.5 $2,150

File attachment

Software development services - hourly - 2023-06-26 through 2023-07-25 17.5 $1,750

File attachment

Software development services - hourly - 2023-07-26 through 2023-08-26 27.0 $2,700

File attachment

Software development services - hourly - 2023-08-27 through 2023-09-25 50.75 $5,075

File attachment

Software development services - hourly - 2023-09-26 through 2023-10-25 31 $3,100

File attachment