Questions about your projects

Questions about a project you built yourself: how to explain it in a minute, defend each choice and name its limits.

1. 01Your project in one minuteEasyAbout 3 min

   Prepare a one minute explanation of one project you built yourself, for example one of the course projects. The interviewer has not seen it.

   Hint

   Five parts in order: question, data, method, result and limit.

   Answer

   State the question in one sentence, then the data, the method, one honest result with a number and one limit. Interviewers check that you can choose what matters and that the numbers hold up to the next question.

   Solution

   Question: what you wanted to find out, in plain words.

   Data: source, period and size, in one sentence.

   Method: the main idea, and the simple baseline you compared against.

   Result: one number with its context, for example a Sharpe ratio after costs on a test period you set aside.

   Limit: one thing the project cannot show. Naming it yourself makes the rest more believable.

   Mark as done
2. 02Is your Sharpe ratio realMediumAbout 6 min

   Your project reports a yearly Sharpe ratio of 0.9 on four years of daily data. The interviewer asks: how sure are you that the strategy has any edge at all? Give a number and explain what you would add to the project.

   Hint

   The standard error of a yearly Sharpe ratio is about one over the square root of the number of years. A slightly better version also accounts for the size of the ratio.

   Answer

   The standard error is about 1/√4 = 0.5, so the ratio is only about 1.8 standard errors from zero. With the refined error sqrt((1 + 0.9^2/2)/4), about 0.59, it is about 1.5. That is weak evidence, so the honest answer is that the edge is not proven yet.

   Solution

   For returns without skill and roughly independent days, the estimated yearly Sharpe ratio has a standard error close to 1/√T with T in years. Here 1/√4 = 0.5, and 0.9/0.5 = 1.8.

   A refinement for a nonzero ratio uses sqrt((1 + SR^2/2)/T) = sqrt(1.405/4), about 0.59, which gives 0.9/0.59, about 1.5.

   Neither clears the usual bar of about 2, and both assume this was the only strategy you tried.

   What to add: a longer or separate test period fixed in advance, results after costs, and the count of variants you tried.

   Interviewers check that you know the error bar of your own headline number.

   Mark as done
3. 03The best of fiftyHardAbout 15 min

   You present a backtest with a yearly Sharpe ratio of 1.0 over 5 years. Under questioning you admit you tried 50 variants and kept the best one. Assume the 50 variants were independent and none had any skill. What Sharpe ratio would you expect from the best of them by luck alone, and how do you defend or repair your result?

   Hint

   Each Sharpe ratio without skill is roughly normal with mean 0 and a standard error of 1/√5. The expected maximum of 50 standard normals is about 2.25.

   Answer

   By luck alone the best of 50 has an expected Sharpe ratio of about 2.25/√5, which is about 1.01. Your 1.0 is what pure chance would deliver, so you cannot defend it as an edge. Repair it with a fresh test period, a correction for the number of trials and an honest report of all 50.

   Solution

   Each zero skill variant has a Sharpe estimate with standard error 1/√5, about 0.447.

   The expected maximum of 50 independent standard normals is about 2.249, computed by integration or simulation.

   So the expected best Sharpe ratio from luck is 2.249 × 0.447, about 1.01, equal to your result.

   In reality the variants are correlated, which lowers the expected maximum. The answer then depends on the effective number of independent trials, which you should estimate.

   Repair: keep the chosen variant fixed and test it on data it has never seen, raise the bar for the number of trials, and show the spread of all 50 results.

   Interviewers check that you count your own trials before they do.

   Mark as done

Compare your own answer with the worked solution.