Full blog post at: https://jamesbachini.com/facebook-advertising-network/

When training the pixel note that Facebook's machine learning algorithms are only as good as the data fed into them.

Make sure your conversions are being fired by people actually interested in the products.  In the past I've setup pixels to fire after 10 seconds just if a visitor stays on the site. This provides little useful data to Facebook and although it looks like you are optimizing faster, it's not effective. 

Video converts better than static content so use this where possible.

Always keep in mind that people are on Facebook to interact with people not businesses. This is why amateurish content and a personalized touch works better than corporate content most of the time. 

Aim to get at least 50 conversion events per ad group per week to grow out of the "learning phase" and give Facebook enough data to optimize with.  Create custom audiences and lookalike audiences based on these conversion events.

Set bid and budget caps to control spending but don't be afraid to test different bids as a higher bid may produce a higher click through rate but lower conversion rate because the ad auctions you are winning are coming from a better demographic. 

This comes from the effect of low bidder getting the auction left overs that companies with more data don't want.  

Scale campaigns both horizontally and vertically by increasing budgets and duplicating the most profitable. Note this will increase ad frequency across the campaigns so it may be worth considering refreshing the ad creatives to avoid banner blindness.

Use https://www.facebook.com/ads/library/ to spy on competitors and see what new creatives leading companies are coming up with. Look for companies that will be doing data-driven marketing at high volumes to get ideas from. 

If you have a customer database, use it. You can convert email addresses or mobile phone numbers into a custom audience for retargeting campaigns.