Quant And Machine Learning Portfolio Construction: Mimicking An Asset Price Series With Nearest Neighbors – Part 3

We present a combination of the ideas from Quant And Machine Learning Portfolio Construction: Mimicking An Asset Price Series With Nearest Neighbors and Quant And Machine Learning Portfolio Construction: Mimicking An Asset Price Series With Nearest Neighbors – Part 2 to construct a portfolio and compute weights to mimic a given time series.

We begin by choosing the SPDR Bloomberg 1-3 Month T-Bill ETF (BIL) to mimic with a stock portfolio. With 150 formation days, we consider stocks with average volume >= 100000 shares. Data is transformed by conversion into a fractional return series, then applying the piecewise aggregate approximation (PAA) with 30 pieces each of length 5 days.

Then we apply the nearest neighbors algorithm using the dynamic time warping (DTW) distance to select 30 stocks that comprise our portfolio.

Using the same DTW distance, we apply the separable natural evolution strategy (SNES) from EvoTorch to find weights, both long only and long/short, that minimize the distance from the portfolio to BIL. This entails selecting weights, simulating the resulting portfolio for 150 days, rebalancing at the end of each quarter, which generates a time series. This time series is then transformed as stated above and the resulting DTW distance to that of BIL is the fitness value that SNES seeks to minimize.
 

Long

Long and Short