Problem

The Google interviewer asks:

How would you forecast airline ticket prices?
How would you correct the Covid impact on the forecast?

Tips

To address this product case question, structure the response in the following manner:

Ask clarifying questions – Get clarity on the forecast horizon of the problem.
Data – What data can you leverage to build the forecasting model?
Methodology – Consider the model design to forecasting the model and the evaluation technique to measure performance.

Solution

[Interviewer] How would you forecast airline ticket prices?

[Candidate] Thank you for the problem. I suppose Google’s search has an airline price prediction that allows users to pick a day to travel based on the price predicted. This is a helpful feature that provides a lot of value to users. Before I start, I have a couple of questions to frame the problem. Do you mind if I ask you?

[Interviewer] Sure thing.

[Candidate] Do I need to forecast the ticket prices on a short-term horizon or long-term horizon? A short-term horizon could be daily prices over the course of one month. A long-term horizon could be daily prices over the course of three or more months.

[Interviewer] That’s a good question. We want to anticipate the daily price six months ahead.

[Candidate] Thank you for the clarification. I have another question. What kind of dataset do I have to build the forecasting model?

[Interviewer] Let’s say you have three years of historical price data and any macroeconomic signals of your choice.

[Candidate] Great thank you. To start, I think it makes the most sense to build a univariate model to get a benchmark, then incorporate macroeconomic signals to improve the forecasting accuracy. Can I start with this process?

[Interviewer] Yes, sounds good.

[Candidate] To start, I need to conduct an exploratory data analysis to apply an appropriate modeling procedure. I would conduct the following analysis that checks for: (1) missingness, (2) outliers, (3) seasonality, (4) trends, and (5) ARMA and MA processes.

[Interviewer] Okay, what’s the next step?

[Candidate] Given that the problem is long-term forecasting, I won’t be able to use LSTM or univariate models such as SARIMA. Such models are designed for short-term models.

Commentary: The candidate is partially correct. There are advanced TS modeling techniques that enable long-term forecasting using short-term models. One such technique is the rolling method.

For instance, suppose that the SARIMA model forecasts 7 days, and the objective is to forecast the next 28 days. Take the output from the first 7 days of the forecast as the input for the next 7 days (days 8 to 14). Continue this procedure until you have a forecast for the entire 28 days.

[Interviewer] Okay, what’s the alternative then?

[Candidate] I could start with a linear model which enables extrapolation. I presume that airline prices are seasonal. For instance, during the summer (in the Northern hemisphere), airline prices would rise as the demand for travel increases. I can build a linear model that regresses time-based signals, including hour, day of the week, week of the month, month, season, and year.

[Interviewer] Okay, how would you build separate predictions depending on locations and airlines?

[Candidate] I could consider two designs. Ultimately, the model performance will determine which one to use. The first design is to build a separate model per location and airline combination. The second design is to build a single model with locations and airlines encoded as categorical features.

[Interviewer] Okay, how would you measure the model performance?

[Candidate] I can evaluate the time-series model using the mean absolute percentage error (MAPE). This metric allows interpretability of the error that’s easy for stakeholders to understand.

[Interviewer] Sounds good. Now, suppose you have macroeconomic signals. What kind of signals would you use?

[Candidate] Can I presume I have signals such as unemployment rate, GDP growth, and population density, and such?

[Interviewer] Yup. Can you elaborate on how those could be helpful?

[Candidate] Well, prices fluctuate depending on the demand. Demand fluctuates depending on the economic conditions. If the economy is bad, people spend less, demand drops, and, ultimately, the price drops.

[Interviewer] Now one final question. How would you accommodate for the Covid impact on the time-series forecast?

[Candidate] That’s an interesting problem. I think leveraging external signals such as lockdown indicators and the Covid confirmed case rates could help build a model that accommodates the Covid impact.

Interviewer Assessment

The candidate is assessed based on three criteria: business/product sense, statistical methodology, and communication. Each dimension contains the following rating: (1) not competent, (2) marginal, (3) adequate, (4) good, (5) superior.

Business and Product Senses – 5

The candidate demonstrated product sense of Google’s Search product – airline price prediction.

ML Methodology – 5

The candidate demonstrated a strong foundation in building a time-series model. Checking for (1) missingness, (2) outliers, (3) seasonality, (4) trends, and (5) ARMA and MA processes are musts in building a TS model.
In general, the candidate understands how to build a long-term forecasting model. Though the candidate isn’t familiar with a rolling forecast technique to build a long-term forecasting model with a model that can’t extrapolate, this issue is quite minor.
The candidate cited reasonable macroeconomic signals for forecasting airline price tickets.
The candidate addressed the solution of adjusting for the Covid impact on the airline price tickets.

Communication – 5

He asked questions to ensure that he understood the problem before diving into a solution.
The candidate explained his approach clearly.

[Google] Forecast Airline Ticket Prices

Problem

Tips

Unlock now by joining.