HOW TO COMPARE YOUR ALGORITHMS
How does one best measure an algorithm’s performance? By all accounts it is not always a simple task.
Algorithm transaction cost analysis (TCA) has been
criticised as a poor science and many believes that
algorithm benchmarks are being abused by gaming
practices that make individual algorithms look good.
“The state of Algo TCA is pretty abysmal” says
Henri Waelbroeck, director of research, Portware.
“It measures performance to a benchmark—but the
question is the performance of what? Orders are of-
ten modified by portfolio managers or merged with
other orders by the trading desk – without taking
the stream of instructions into account TCA be-
The solution, says Waelbroeck, is to construct a re-
alistic baseline price for blocks by tracking the stream
of portfolio manager instructions and order merging
requirements, then allocating results—the realised
price against the baseline—back to portfolio manager
orders to account for costs for each portfolio.