To retrain or not to retrain, enhancing results by simple methods

Keywords: FX, Currencies, Indices, MLP, Training, systematic trading, Investing, algorithms, GDP, Sharpe, Neural networks, machine learning, deep learning.

Retraining is simply the process of training the whole dataset when a new data is added. Incorporating the latest actual value to forecast tomorrow's value makes more sense than choosing an out-of-sample period and assuming that the relationship doesn't change. The study will compare the following scenarios (OOS = Out-of-sample):

• A 3000-long dataset out-of-sample period without retraining.
• A 3000-long dataset out-of-sample period with retraining assuming same initial conditions*.

* No grid searching throughout the retraining, only done at the initial phase to find optimal hyperparameters.

The performance measures used to analyze the results will be:

1. Sharpe ratio assuming a 3% annualized risk-free rate (Thus, roughly 1.5% for 6 months).
2. Gross return over the period of 3000 trading hours.
3. Accuracy, also known as the hit ratio.
4. Relativity to the benchmark (outperformance or underperformance).
5. The percentage of long positions to measure the bias of the model.

The test asset will be an interesting liquid FX pair, the GBPUSD (10-week volatility = 0.75%). Unlike the EURUSD (10-week volatility = 0.53%), the GBPUSD is not boring and actually moves from time to time which is why it can present an interesting opportunity. The training data will be 15,100 and the test data is 3,000 thus providing a 20/80 split ratio. The following table summarizes the results:

Conclusion

Retraining seems to deliver some added-value. The accuracy and return have improved but in no way does that mean retraining is better than assuming a stable relationship over time (Which in this case was 6 months) for the following reasons:

• Only one test asset has been examined.
• Only one period of time has been tested.
• The %long has not changed by that much.

Retraining can be further investigated and enhanced through many other ways, but for now, its intuition is appealing and calls for more thorough research.