
Learning how to show work on machine learning problems means deciding which lines a reader needs to see. If you write only “Model A wins,” nobody can tell which rule you used or whether you applied it correctly. If you spell out every addition, the key decision gets buried. A good handwritten solution leaves a short, checkable path from the information in the prompt to the answer.
Start by finding the decision the problem asks you to make. If it asks which model has lower mean squared error, you need to compute that error for both models and compare the results. You do not need to explain how either model was trained if the problem already gives you its predictions. That distinction tells you where your solution should begin.
Try this example: Two models, A and B, predict how many stickers three students will use to decorate their notebooks. In row order, the actual counts are 4, 6, and 8. Model A predicts 5, 5, and 9. Model B predicts 4, 6, and 6. Which model has the lower mean squared error on these three examples?
The two models make different kinds of mistakes. Model A is a little off on every row. Model B gets two rows exactly right but misses the third by more. Counting exact matches might make B look better, but the problem asks you to compare squared errors, not count exact matches. Name that rule before you do the arithmetic.
On paper, one sentence can establish what the symbols mean: “Let i be the row number, yi the actual count in row i, and pi the prediction from the model being evaluated in that row.” Then write the rule:
MSE = [(p1 − y1)2 + (p2 − y2)2 + (p3 − y3)2] / 3.
The denominator is 3 because each model makes one prediction for each of the three rows. Use the same definition for both models and label each calculation. There is no need for a second set of symbols for B if the label makes clear that its predictions go into the same formula.
If the problem already defines its own symbols, keep them unless you have a reason to switch. A short note such as “All lists below follow the row order in the prompt” can do more for readability than introducing a new variable. The goal is to make it clear which prediction goes with which actual value.
For this problem, the useful intermediate values are the signed errors, also called residuals here, and the squared errors that go into the average. Write one subtraction in full to establish your convention: “For model A in row 2, the residual is 5 − 6 = −1.” Then show the same step for the remaining rows in a compact form.
| Model | Residuals in row order | Squared errors in row order | Mean squared error |
|---|---|---|---|
| A | 1, −1, 1 | 1, 1, 1 | (1 + 1 + 1) / 3 = 1 |
| B | 0, 0, −2 | 0, 0, 4 | (0 + 0 + 4) / 3 = 4/3 |
Two labeled lines can replace the table if you are short on space. Either way, keep the step that turns B's last residual of −2 into a squared error of 4. That value drives the comparison. You can skip writing that 0 + 0 equals 0 or converting 4/3 to a decimal. Neither step helps the reader check how you applied the rule.
Keep the order of operations visible, too. Averaging the signed errors and then squaring the result would let positive and negative errors cancel before squaring. That is a different calculation. The squared error column makes clear that you squared each error before averaging.
Write the conclusion as a sentence: “Model A has the lower mean squared error on these three examples because 1 < 4/3.” The inequality makes the comparison explicit. No rounded values are needed, and the phrase “on these three examples” keeps the claim within what you actually checked.
If the prompt also asks why B lost despite getting two predictions exactly right, point to the numbers: B's one squared error is 4, while A's three squared errors add up to 3. Both totals are divided by 3. That explains the result without repeating the entire table. It does not prove A will make better predictions on new examples or score better under another rule.
Cover your scratch work and read only the solution you plan to hand in. Can you tell what the symbols mean, where the two scores came from, and why your final sentence follows? If a score seems to appear from nowhere, add the line that produced it. If the line is already there, make its label easier to spot instead of copying the calculation.
Try the same check on a problem from the training and practice problems. If you plan to submit handwritten work for the Qualification Round, the competition rounds page says that a clear picture of your notes is an option. In the photo, check that minus signs and superscripts are easy to distinguish and that the final line is not cut off.
Do I need to show every subtraction?
Not if one example makes the method clear and you show the resulting values in a labeled list or table. Keep extra detail when a step is easy to misread or changes the result.
Can I leave my answer as a fraction?
Yes, unless the problem asks for a decimal or a particular precision. An exact fraction can make a comparison clearer than a rounded decimal.
Is a table enough for my final answer?
Use the table for calculations, then add a sentence that answers the question directly. For a model comparison, name the model, the metric, and the examples the comparison applies to.
About the International Machine Learning Competition: an online machine learning competition for high school students from around the world. Qualification Round 2026: submissions are open until Sunday, 13 December 2026. How the rounds work