Key points of the talk :

  • Who’s doing the work?

  • Financial economists – Develop Equilibrium models, inventory control models, asymmetric information models. Some strategic equilibrium in the background

  • High Frequency Econometricians – Less concerned with issues about strategic equilibrium. Working on predicting realized volatility. When you hear the word, “ Microstructure noise”, you can be reasonably certain that it is a HFT econometrician

  • Mathematicians – NYU Algo trading conference had about 300 mathematicians who had a fair enough IT background. These are professionals working in private sector who care less about equilibrium model, and are working on developing trading strategies.

  • Increasingly mathematicians are the market participants. Dealers/Specialists/Block trading were the key players in the yesteryears. Today it is the ATS, ECNs, dark pools are the places where 2 out of 3 trades that take place away from the primary exchanges.

  • Most of the academic microstructure is devoted to the equity markets.

  • Classic Microstructure models were dealer oriented. They were usually inventory control theories and asymmetric information models. It lead to spread decomposition models, price impact models, etc. Most models followed from dealer/specialist paradigm.

  • If the data had some problem, one could have concentrated on NYSE or NASDAQ as that was the primary mode of exchange. TAQ was a plausible public analysis dataset. No longer can one do so as majority of the trades happen off these exchanges.

  • Subscriber feeds rather than consolidated feeds are being used for quant modeling

  • World of continuous double auction - Difficult to get practical models

  • Estimation of price impact functions (Cannot trust the sign for the trade)

  • How long one does have to wait before a cancellation occurs? Statistical tools like duration analysis can be used.

  • Seasonality in intraday data - 1 second periodicity in the data

  • Three Case studies Takeaway: Order submission and cancellations increase phenomenally in a time window but the executions do not go up.

  • For episodic periodicity, one can use wavelets.

  • Individual limit order fills are very low ~ 1%

  • Liquidity demander’s price impact was the focus of modeling. By looking at the data, one could make out who was the dealer, who was the taker. Based on bid and ask process, in the pre-decimalization years, one could confidently put a sign on the direction of the trade. In the HFT scenario, one could get completely wrong sign on the trade.

  • We cannot trust the clean dichotomy between liquidity demanders and liquidity suppliers. Dealer – customer distinction is a lot more fluid.