|Date:||01 November 2019|
I am interested in forecasting in financial markets. Several years of forecasting experimentation. What have we learned so far?
- Dataset matter a lot.
- Methods can enhance forecasting ability.
- Specifications affect the performance
There has been progress but there is room for improvement. Significant improvements can come
after answering the following generic topics:
- Forecasting Financial Time Series - Can we forecast returns?
- Dimensionality Reduction Techniques - Does this massive information makes sense?
- Alternative Predictors - Should we base the forecasting process only on the information contained
- Forecasts’ Evaluation - How close is a candidate model to the actual data?
- Signal Decomposition - Can we determine which component of a variable is more relevant?
Ideas for thesis topics:
Expanding the work of Stock and Watson (1998, 2002) and Ng and McCracken (2016), the first proposed topic aims at generating super-datasets for the largest economies.
Following Rossi and Sekhposyan (2011), providing good point forecasts is not enough, we should
be able to report the uncertainty around the forecasts.
Faria and Verona (2018) show that pooling the forecasts of frequency decomposition signals can improve the forecasts. Can simple machine learning techniques allocate the optimal weight on the forecast of each signals?
Rapach and Strauss (2012) point out that sole models are making the a priori selection of the right model extremely difficult. Hence, they propose an amalgamation of forecasts that uses information from all methods.
The impact of skewness a Kurtosis on Realized Volatility forecasting, as an extension of the work of Degiannakis and Filis (2018).
Taylor rules are extremely popular in the literature. Stepping on the work of Amat, Michalski and Stolz (2018), the last topic proposes the calculation of Taylor rules coefficients with machine learning techniques.