Join waitlist
The books... not free though
10 Strategier Som Slår BörsenSvenska10 Strategier Som Slår Börsen10 Strategies That Beat the MarketEnglish10 Strategies That Beat the MarketProRealTime Coding Made SimpleEnglishProRealTime Coding Made Simple
600+StatisticsMarket statistics and edges1,327StrategiesBacktested strategy library18Research papersThe study behind the algos301DocsProBuilder coding reference27Algo development programCourses and lessons24Everything here feeds the 24 algosSee the algos

Stats › Backtesting and evidence › Machine learning

The best machine learning forecast explains 0.40% of monthly return variance

Elastic net
0.11%
Least squares, 3 predictors
0.16%
Generalised linear model
0.19%
Principal components
0.26%
Partial least squares
0.27%
Random forest
0.33%
Boosted trees
0.34%
Neural network, three layers
0.40%

Monthly out of sample R squared for individual stock return forecasts, all stocks, out of sample window 1987 to 2016 with annual refitting. The benchmark is a forecast of zero rather than the historical mean, which mechanically raises every figure. Unregularised least squares on all 920 predictors is omitted here and scores minus 3.46%.

Neural network forecasts produced a monthly out of sample R squared of 0.40% and a long short Sharpe ratio of 1.35 value weighted and 2.45 equal weighted. Neither of the two most cited studies reports any transaction cost adjustment at all.

0.40%Best monthly out of sample R squared across all methods
0.16%The same for a three predictor linear model
-3.46%The same for unregularised least squares on 920 predictors
NonePublished cost adjusted Sharpe ratios located

Members only

The rest of this page is for members

Below this point there are 6 sections and 5 named sources, roughly 900 words of it. Every figure carries the source it came from and the date the data is from.

You can keep browsing every statistic in the library for free. The intro and the summary are always open.