Stake on Signals, vs. Trade Directly

As noticed in another thread: what one attempts to predict for Signals, is directly applicable to managing one’s own portfolio. Of course, because that’s what it’s designed for!

So I’m introducing this topic, to sound you all out about relative advantages (how staking is better than trading) and disadvantages (how staking is worse), and perhaps some fundamental differences I’ve not though of.


I am primarily Numeraire investor and I think of the gains as an additional edge. If you trade in $ your money are backed by depreciating asset.

with staking, you have none of the following:

transaction costs, borrow fees, margin call risk, prime broker fees, etc.

Sure, you could potentially deploy your model on the live stock market, but I hope you have enough money to buy all the positions your model calls for! Imagine your top long pick is BRK.A!


Would be a cool experiment to create model and stake it if corr/mmc looks good and then using a paper trade account and buying say top 10 or 20 highest stock signals sorted out Round start and selling them all round end for several months and seeing what happens

I feel safe to stake on signals. But looking at the pattern of tournament, not sure how long the payout rate can be sustained. At the end, everything we learned from numerai, should be reapplyed to the live trading environment.

I fully agree.

Every thing is in the cards for a much lower average payout in main tournament (best models will probably get the same), and a higher payout in Signals. Data Scientists will only start staking big in Signal if they think they can make at least the same return over time as in the tournament.

I’ve been doing this for 4 weeks so far on the demo side of my UK spread bet account. Choosing just a few stocks with the highest expected returns to see what the ride might be like. Overall it looks promising but would need a proper backtest, and promising results can fall apart when switching to live, particularly if there’s any discretionary element, which ideally is best avoided.

My personal experience has been, it’s ok to use something like the Signals pipeline as an initial filter. If you’re trading small baskets, you could add sanity checks with a secondary trend or regression forecast (I use fbprophet), then optimize your allocation with pyport.

1 Like