Case study: It was a one-hour interview with two documents to read. The first provided background and a short introduction to the electricity market in power trading.
The market participants included ordinary consumers, utility companies, electricity producers, and electricity traders. The basic assumptions were:
Electricity prices stayed constant within a day, and ordinary consumers did not change their consumption in response to price fluctuations.
Electricity could not be stored, so excess production was simply wasted.
Utility companies had to prioritize preventing blackouts and would try as hard as possible to avoid insufficient supply.
Traders could not take physical delivery and had to close their positions before delivery.
Part 1
The interviewer and I first discussed this background. The main questions were how changes in electricity demand might affect each participant under this model and, if we forecast electricity demand, which metrics each participant would care about. Because this was an interview, we focused mostly on the perspectives of traders and energy companies. I did not have a power-trading background, so I reasoned mainly from the provided material.
For traders, I thought there were more financial instruments they could use and different possible prediction targets. We could forecast volatility and consider building an options portfolio. If we forecast the direction of prices, we would need to choose a direction, an entry time, and a time to close the position, because physical delivery was not allowed.
From the company's perspective, I said it would focus on the standard deviation of the forecast. It would want to minimize variance and uncertainty as much as possible. That would help it avoid blackouts without wasting electricity through excess production.
Part 2
Suppose there were a black-box machine that predicted prices for the following week and cost 75,000 per year. How would I evaluate the machine and decide whether to subscribe?
We also had a lot of public data, such as weather and policy-related data. Everything was daily. If I received the machine's past five years of forecasts alongside the true values, how would I use them?
I planned to update this part later, but my rough idea was to use linear regression as a benchmark and then measure how much more accurate the machine was.
The final situation was that we had bought the forecast, but its accuracy declined for five consecutive days. How would I diagnose whether we should stop using it or whether an explainable event caused the performance drop?
Because I am not familiar with energy trading, I was curious how people with more experience would answer these questions.
Discussion
Loading comments…