I am finding more and more that the HFT papers have some of the best insights into machine learning for finance.

I often get asked what are some good features to include are.

The following paper: Execution Strategies in Equities Markets by M.G. Sotiropoulos is a great example:

Trade autocorrelation: trade signs are correlated within short timescales due to algorithmic slicing of large client orders; this could lead to predictable trade direction.

Order imbalance: the limit order book may be too heavy on one side; this could lead to predictable price movement over the next time period.

Momentum/reversion: the price path exhibits a strong trend; momentum traders bet on the trend persisting and reversion traders bet on its reversal.

Relative value: the traded asset is cointegrated with a sector index or another asset; this leads to predictability in the reversal of the spread between the traded and the reference asset (pairs trading).

Volume clustering: a recent spike in trading volume may be a leading indicator of more volume spikes in the short term.

Unexpected news: the market responds strongly to unscheduled news about a company or product; this may lead to a predictable increase in trading volume or even a predictable price trend.

Informed trading: the presence of agents with private information may be inferred from the order flow using the probability of informed trading indicators PIN and VPIN as in Easley et al (2011); Venue toxicity can affect the trading decisions of uninformed agents

Venue liquidity: high activity in a certain trading venue (lit or dark) may be used to adjust the probability of getting filled across venues and reallocate quantities at the smart order router level.

You can read the rest via Google scholar: https://scholar.google.com/scholar?hl=en&as_sdt=0%2C5&q=execution+strategies+in+equities+markets+Michael+G+Sotiropoulos&btnG=