IB Math IA ideas: statistics and probability

Statistics and probability is the largest pool of IA ideas, because data is easy to find or collect. It is also where the subject reports see the most weak work: two variables and a correlation, or a test applied without checking its conditions.

By Pietro Meloni, PhD · Updated on

42 of 42 ideas

Criteria referred to in each idea: A Presentation, B Mathematical communication, C Personal engagement, D Reflection, E Use of mathematics. What each criterion rewards.

Topic 4, statistics and probability: 42 ideas

Monty Hall with n doors and a forgetful host

Aim. How does the probability of winning by switching change when there are n doors, the host opens k losing doors, and the host may or may not know where the prize is?

  • AA and AI · SL and HL
  • Easy to start
  • Simulation
  • Overdone: on 4 sites

My take. Overdone in its basic form. It only works with the forgetful host case and the general n and k formula, which is where the real mathematics is.

Mathematics, first steps and where marks are lost+
Mathematics
Conditional probability; tree diagrams; law of total probability; Bayes' theorem for the case when the host opens a door at random; generalising to a formula in n and k; verification by a simulation with a large number of trials. Beyond syllabus: expected values of strategies over repeated games.
How to start
State the rules for the classic game and compute the switching probability with a tree diagram. Then write a short simulation of 100,000 games for n = 3, then for n = 5 with k = 1, 2, 3, and compare it with your formula.
SL and HL
SL builds trees and derives the general formula for n and k. HL adds Bayes' theorem in the case of a random host and the discussion of how the information given by the host's action changes the result.
Where marks are lost
C: a very familiar problem with no personal angle. E: staying on the 3-door case that appears in every textbook. D: no comment on the sensitivity of the result to unstated assumptions about the host. B: ambiguous rules.
Data
Pure theory with a simulation for checking, so no data is needed, but the rules must be stated with precision.

Queueing at a real junction: arrivals, service and green time

Aim. Can a Poisson arrival model and a fixed discharge rate predict queue lengths at a local traffic light, and what green-time split minimises the average wait?

  • AI · HL topic
  • Needs care
  • Your own data
  • Common: on 3 sites

My take. With a real junction and a testable model this is strong. The twist is a junction you use every day.

Mathematics, first steps and where marks are lost+
Mathematics
Poisson distribution for arrivals, chi-squared goodness of fit test, expected queue length from a deterministic queue model, optimisation by differentiation of total delay with respect to green time. Beyond syllabus: the M/M/1 queue formula L = ρ/(1 − ρ) and its breakdown as ρ tends to 1.
How to start
Film a junction for 30 minutes and count cars arriving in every 10-second interval. Test the counts against a Poisson distribution, then measure how many cars leave per second of green.
SL and HL
The Poisson distribution is AI HL. HL students fit and test the Poisson model, build a delay function and optimise the green split by differentiation, then compare with the M/M/1 result. An SL student could model arrivals per interval with a binomial distribution instead.
Where marks are lost
E: listing formulae without using them on real data. B: undefined symbols such as ρ. C: no observation made by the student. D: no discussion of rush-hour counts breaking the Poisson assumption.
Data
Counting cars takes a few hours of recording and tidying, but produces authentic data.

Bluffing frequency in a simplified poker game

Aim. In a one-card betting game, what mix of bluffing and calling makes a player's expected winnings unexploitable?

  • AA and AI · SL and HL
  • Needs care
  • Pure maths, no data
  • Common: on 2 sites

My take. The full poker version is too broad and heavily used. A tiny game solved exactly, then played by classmates against the prediction, turns it into a good exploration.

Mathematics, first steps and where marks are lost+
Mathematics
Combinatorics for hand counts, conditional probability, expected value, payoff matrices, solving linear equations for the indifference condition, mixed strategy. Beyond syllabus: Nash equilibrium in a simplified Kuhn poker.
How to start
Set up a three-card game with one bet size. List every deal and every action, then find the expected payoff of each pure strategy pair in a table.
SL and HL
SL works with the payoff table and expected values. HL solves the equilibrium mixing probabilities with algebra and may generalise to a variable bet size.
Where marks are lost
E: real Texas Hold'em is too big, so the maths becomes descriptive. B: game rules poorly explained. D: failure to test the model against play.
Data
Mathematics is controllable only if the game is cut down; testing against real play needs self-collected hands.

Buffon's needle as a way of finding π

Aim. How many needle drops are needed to estimate π to a given accuracy, and does the choice of needle length relative to line spacing make the estimate more efficient?

  • AA and AI · SL and HL
  • Easy to start
  • Simulation
  • Common: on 2 sites

My take. Popular, so it needs the efficiency question to stand out. Adding a real drop of 500 matchsticks makes it your own.

Mathematics, first steps and where marks are lost+
Mathematics
Uniform distributions for the centre distance and angle, P = 2l/(πd) by integrating over the angle, expectation and variance of a binomial count, standard error of the π estimate, simulation, dependence of variance on l/d. Beyond syllabus: using the delta method to approximate the variance of the π estimate.
How to start
Derive P for l ≤ d, then write a spreadsheet or short program that drops 1000 random needles and estimates π; repeat 100 times to see the spread.
SL and HL
SL does the geometric derivation with a simple integral and a simulation. HL or top work examines variance against l/d, shows the best ratio and quantifies the convergence, including a confidence interval.
Where marks are lost
E: a simulation without a derivation or error analysis is too thin. C: a very common topic, so no physical experiment or personal angle. B: unclear model of what a random drop means.
Data
Simulation is easy; a real throw of needles is optional but slow.

Can standing height predict a volleyball player's spike reach?

Aim. How well does a linear model based on standing height predict the spike reach of professional volleyball players, and does the fit differ between outside hitters and middle blockers?

  • AI · SL and HL
  • Needs care
  • Public data
  • Seen too often in subject reports
  • Common: on 2 sites

My take. Two variables and a correlation is named in the subject reports as a common and usually weak format, so go beyond r. The residual analysis and position comparison give it substance, and measuring your own club team as an extra check makes it personal.

Mathematics, first steps and where marks are lost+
Mathematics
Scatter diagram, Pearson's correlation coefficient, least-squares regression line of y on x, the coefficient of determination, residual analysis with a plot of residuals against fitted values, interpolation versus extrapolation, and a two-sample t-test (AI) on the residuals of two positions, with a check of the normality assumption. Beyond syllabus: a confidence interval for the slope, or separate lines for each position with a comparison of gradients.
How to start
Pick one recent international tournament, such as a Nations League or Olympic squad list, and copy height, spike reach and position for every player into a spreadsheet. Record the source of each row. Draw the scatter plot first and compute r and the regression line with a graphing calculator or spreadsheet before doing any testing.
SL and HL
An SL student can complete the scatter plot, r, the regression line and a careful interpretation of gradient and intercept, with a short residual discussion. HL depth comes from arguing why the test conditions hold, comparing position-specific models, and showing by hand why the least-squares line minimises the sum of squared residuals for a small data set.
Where marks are lost
E: stopping at r and a line of best fit with no test of the model's assumptions; treating a high r as proof of cause. B: unexplained calculator output and no definitions of spike reach. C: generic data with no personal link, such as no connection to the student's own team. D: no comment on the bias of using elite players only, or on measurement differences between sources.
Data
Data are public but need cleaning and consistent definitions, and a sample under about 30 per group weakens the t-test.

Cooperation in the repeated prisoner's dilemma

Aim. How does the probability that a game continues affect which strategy earns most in a repeated prisoner's dilemma?

  • AA and AI · SL and HL
  • Easy to start
  • Simulation
  • Common: on 2 sites

My take. The one-shot version is weak and known to moderators. The repeated version with a derived critical probability is a solid exploration.

Mathematics, first steps and where marks are lost+
Mathematics
Payoff matrices, expected value, geometric series for an infinite repeated game with continuation probability p, comparing sums of series, inequalities to find the critical p. Beyond syllabus: Markov chains for strategies with memory.
How to start
Choose payoffs such as 3, 5, 1, 0. Compute by hand the total for always defect against tit-for-tat over n rounds. Then find the p above which tit-for-tat beats defection.
SL and HL
SL sums geometric series for a few strategy pairs. HL builds a round robin tournament and analyses strategies with memory using transition matrices.
Where marks are lost
E: one-shot game only, which is just reading a table. B: strategy definitions ambiguous. C: retelling Axelrod's tournament without original work.
Data
Simulation is simple; the series analysis gives the real mathematics.

Do leading digits expose invented numbers?

Aim. Do leading digits in several real datasets match the logarithmic distribution of Benford's Law, and can the same test separate them from numbers I invent or that classmates invent?

  • AI · SL and HL
  • Easy to start
  • Public data
  • Common: on 2 sites

My take. Very popular, so a routine version will look familiar. Make it distinctive with an experiment where classmates fabricate data and you measure how well the test catches them.

Mathematics, first steps and where marks are lost+
Mathematics
Probability of a leading digit as log10(1 + 1/d), properties of logarithms, chi-squared goodness of fit test with degrees of freedom and p value, comparison of expected and observed frequencies. Beyond syllabus: why scale invariance forces the logarithmic law.
How to start
Take populations of countries and a set of river lengths or share prices, tally first digits in a spreadsheet, then compute the expected counts from the formula and a first chi-squared value.
SL and HL
SL derives the expected proportions, runs the chi-squared test on two or three datasets and compares with invented data. HL or top SL adds a scale-invariance or geometric-series argument for why the law arises, and examines why some datasets fail, such as heights.
Where marks are lost
E: applying chi-squared with expected counts below 5 or with a huge sample that rejects everything. B: not defining hypotheses. D: claiming the test proves fraud. C: a purely descriptive account of the law.
Data
Plenty of open data; the care needed is in choosing sets that span several orders of magnitude.

Do rock-paper-scissors players follow a predictable Markov chain?

Aim. How well does a first-order Markov chain, fitted to recorded rock-paper-scissors games, predict a player's next throw, and does a strategy built on that chain beat a player choosing at random?

  • AI · SL and HL
  • Needs care
  • Your own data
  • Common: on 2 sites

My take. A sound choice that is only moderately overused, provided the test of the Markov assumption is real and not decorative. The personal twist is to play against one specific opponent, such as a sibling or a friend, and see whether the chain exposes their habits, then whether the habit disappears once they know.

Mathematics, first steps and where marks are lost+
Mathematics
Relative frequencies estimated into a 3 by 3 transition matrix; conditional probability P(next | previous); expected value of a strategy from its payoff table; a chi-squared test of independence between consecutive throws; matrix powers to find the long-run distribution; HL: the steady state as an eigenvector with eigenvalue 1, and the speed of convergence from the second eigenvalue. Beyond syllabus: comparing first-order with second-order chains by counting parameters.
How to start
Recruit 15 to 20 classmates, ask each for 60 to 100 throws against a computer or a hidden partner, and log each sequence in a spreadsheet. Then count the nine previous-to-next pairs for one player and turn the counts into a transition matrix.
SL and HL
SL builds the matrix from counts, computes conditional probabilities, and tests whether the win rate of a follow-the-pattern strategy exceeds one third using simulation or a simple test. HL derives the steady state as an eigenvector, links it to the fairness of the game, and compares the fitted chain with the null model of independent uniform throws, with a discussion of overfitting when only 60 throws feed nine parameters.
Where marks are lost
E: matrix and steady state done mechanically with no test of whether the Markov assumption fits; C: generic survey with no personal angle; D: no discussion of small samples, players who adapt when they lose, or the difference between a fitted pattern and a real one; B: notation for state, row and column convention left unclear.
Data
Collecting enough throws per player is the bottleneck, since nine matrix entries need several hundred throws in total for stable estimates, though pooling players or using an online game log solves it.

How wrong can a poll be? Sampling error against actual results

Aim. How often did the published margin of error of UK opinion polls contain the actual election vote share, and what does this suggest about the assumptions behind a normal model?

  • AA and AI · SL and HL
  • Needs care
  • Public data
  • Common: on 2 sites

My take. Election poll ideas are common, so focus on the coverage question instead of predicting the next winner, and pick elections you have a reason to care about.

Mathematics, first steps and where marks are lost+
Mathematics
Sampling distribution of a proportion, normal approximation, 95% confidence interval p ± 1.96√(p(1 − p)/n), comparison with actual results, mean error and root mean squared error, simulation of repeated polls. HL: hypothesis test or chi-squared on the frequency of misses, and the effect of design effects or house bias.
How to start
Download the final polls before three or four past UK general elections from a public archive, with sample sizes, and compute each margin of error. Then tabulate whether the true party share lay inside the interval.
SL and HL
SL: compute intervals, count coverage and simulate. HL: test whether the miss rate exceeds 5% and model systematic bias using regression of error against date or pollster.
Where marks are lost
E: simply drawing a regression line through polls without justification. C: reflecting no personal interest, and stale politics. D: not discussing non-sampling error such as late swing.
Data
Data is public but pollster methodology and sample sizes have to be cleaned and matched to results.

Income and life expectancy across countries

Aim. Is life expectancy better modelled as linear in GDP per capita or in its logarithm, and how has that relationship changed between 1990 and the latest year?

  • AA and AI · SL and HL
  • Easy to start
  • Public data
  • Seen too often in subject reports
  • Common: on 2 sites

My take. Two variables and a correlation is named in the subject reports as a common and usually weak format. Choose it only for the log transformation and the change over time, or link it to a country you know.

Mathematics, first steps and where marks are lost+
Mathematics
Scatter plots, Pearson and Spearman correlation, least-squares regression, a log transformation of GDP, residual plots, coefficient of determination, comparison of the fits. Beyond syllabus: explaining why diminishing returns suggest a logarithmic model.
How to start
Download GDP per capita and life expectancy for all countries for two years from Gapminder or a World Bank table, remove missing entries, and draw the raw scatter plot.
SL and HL
SL fits linear and log models and compares r and residuals. HL adds Spearman's rank correlation (AI), residual analysis and a non-linear fit; the extra depth comes from comparing years.
Where marks are lost
E: only computing r on raw data is too shallow. C: a very common topic, so choosing a niche question is essential. D: treating correlation as causation, and ignoring outliers or confounders such as health spending.
Data
Clean public data exists, but cleaning missing countries needs a stated rule.

Is the spreadsheet's random number generator really uniform?

Aim. To what extent do the values from a spreadsheet's random function behave as independent draws from a uniform distribution on (0,1)?

  • AI · SL and HL
  • Easy to start
  • Simulation
  • Common: on 2 sites

My take. Neat and safe. The twist that lifts it is building your own poor generator and showing your tests catch it, a good way to judge test power.

Mathematics, first steps and where marks are lost+
Mathematics
Uniform distribution properties, frequency tables with equal bins, expected frequencies, chi-squared goodness-of-fit, correlation between consecutive values, a runs test or a chi-squared test on pairs. Beyond syllabus: a Kolmogorov-Smirnov statistic.
How to start
Generate 5000 values, bin them into ten classes and calculate the chi-squared statistic and p-value. Then plot each value against the next and test the pair counts.
SL and HL
SL tests uniformity and correlation between consecutive values. HL adds a test on pairs with a two-dimensional chi-squared, a comparison with a simple linear congruential generator built by hand using modular arithmetic, and a power analysis.
Where marks are lost
E: one test on one sample with no consideration of repeated runs. D: treating 'p above 0.05' as proof of randomness. B: unclear hypotheses.
Data
Data is instant, but repeated runs and the correct degrees of freedom need care.

Minimum-variance weights for a three-asset portfolio

Aim. For three stocks chosen from different sectors, which weights minimise portfolio variance for a required expected return, and how far does the fitted frontier predict the next year's actual results?

  • AA and AI · SL and HL
  • Easy to start
  • Public data
  • Common: on 2 sites

My take. Common, but it works if the optimisation is done by hand and validated on later data. Choose assets tied to your own interest and be honest that the model often fails out of sample.

Mathematics, first steps and where marks are lost+
Mathematics
Expected value and variance of a weighted sum of returns, which brings in covariance (beyond the syllabus, but a short step from variance); for two assets the variance is a quadratic in the weight, minimised by differentiation or completing the square; for three assets, substitute w₃ = 1 − w₁ − w₂ and solve the linear system from setting both derivatives to zero; out-of-sample comparison. AI HL students can write the variance in matrix form. Beyond syllabus: Lagrange multipliers and the efficient frontier.
How to start
Download five years of monthly closing prices for three stocks from Yahoo Finance, convert to monthly returns, compute means and the covariance matrix, then find the minimum-variance weights for two assets by hand before moving to three.
SL and HL
An SL student can do two assets with the variance as a quadratic in w, then compare a few chosen three-asset weightings. HL depth comes from the three-asset system, the matrix formulation and testing the weights on a later period.
Where marks are lost
E: using Excel Solver as a black box; B: matrix notation introduced without explanation; D: no comment on the assumption that past covariances persist; C: generic choice of stocks with no reason.
Data
Prices are publicly available; cleaning is minor, though returns must be computed consistently.

Mixed strategies in penalty kicks: theory against real shots

Aim. Which left/right mixture leaves a penalty taker with no exploitable weakness, and how close do professional kickers and keepers come to it in real shootout data?

  • AA and AI · SL and HL
  • Needs care
  • Public data
  • Common: on 2 sites

My take. A sound idea if the data is tested honestly. The twist is to use your own club's penalties or a single tournament, and to report where the equilibrium fails.

Mathematics, first steps and where marks are lost+
Mathematics
Payoff matrices of scoring probabilities, expected value, indifference conditions that give a mixed-strategy equilibrium by solving linear equations, then a chi-squared goodness-of-fit test (AI) comparing observed direction frequencies with the predicted mix. Beyond syllabus: the minimax idea behind Nash equilibrium, and adding a third option (centre).
How to start
Take penalty records for one league or tournament from a public football dataset, tabulate kicker side (natural or weaker foot) against keeper dive direction, and compute the four scoring rates. Then solve for the mix that makes the keeper indifferent.
SL and HL
SL solves the 2 by 2 game with expected value and simultaneous linear equations and compares it with observed proportions. HL adds a sensitivity analysis of the equilibrium to sampling error, AI HL confidence intervals on the scoring rates, and a 3 by 3 extension including the centre.
Where marks are lost
E: stopping at a textbook 2x2 game with no testing of the data. C: generic football enthusiasm with no personal angle. D: not questioning the assumption that kickers choose simultaneously and rationally. B: undefined game theory vocabulary.
Data
Penalty data exists online but cell counts can be small, and direction has to be coded as natural side versus weaker side consistently.

Shared birthdays: uniform theory against real birth data

Aim. For a group of n people, how does the probability of a shared birthday change when real, uneven birth frequencies replace the uniform assumption, and what group size gives an even chance in each model?

  • AA and AI · SL and HL
  • Easy to start
  • Public data
  • Seen too often in subject reports
  • Common: on 2 sites

My take. The birthday problem is named in the subject reports among topics seen too often, so only choose it if real birth data and the comparison of models are the centre. Testing your own year group or school against the prediction makes it personal.

Mathematics, first steps and where marks are lost+
Mathematics
Complementary events and the product of conditional probabilities, solving 1 − 365!/((365−n)!·365n) = 0.5 by table or logarithms, Poisson-style approximation using exp(−n(n−1)/730), a weighted model with daily probabilities pi and the sum of pi squared as a collision measure, Monte Carlo check. Beyond syllabus: a short argument that uniform probabilities minimise the collision chance (Cauchy-Schwarz).
How to start
Download a public table of births per calendar day (for instance national birth statistics from a statistical office over several years), convert to relative frequencies, and first compute the uniform answer for n = 23 on a spreadsheet before touching the real data.
SL and HL
SL derives the uniform result, finds the threshold by trial and compares it with a simulation. HL or top-band work adds the weighted model, the approximation by exponentials and a proof or numerical demonstration that any unevenness raises the match probability.
Where marks are lost
C: the topic is the standard one, so personal engagement depends on using real data and a twist of one's own. E: staying with the uniform calculation only is too thin. B: unexplained use of factorials on a calculator. D: no comment on ignored leap days and clustering of births.
Data
Births by day are publicly available, though leap day and rounding need a stated decision.

Sleep hours and school results: what does a survey show?

Aim. How strongly is self-reported sleep duration associated with a defined measure of school performance in my school, and do a numerical and a banded analysis reach the same conclusion?

  • AI · SL and HL
  • Needs care
  • Your own data
  • Seen too often in subject reports
  • Common: on 2 sites

My take. Two survey variables and a correlation is named in the subject reports as a common and usually weak format. It is workable only with your own data, a strict definition of the performance measure, and honesty about association and cause.

Mathematics, first steps and where marks are lost+
Mathematics
Scatter plot, Pearson and Spearman correlation, least-squares regression, grouping into bands and a chi-squared test for independence with expected counts and degrees of freedom, a check that expected counts exceed 5, a discussion of the effect of banding choices.
How to start
Design a short anonymous questionnaire with one question on average sleep and one on a clearly defined result (for example last term's mean mark banded), get school permission and collect at least 60 responses.
SL and HL
SL does correlation, regression and one chi-squared test. HL or top work tries different bands, checks sensitivity, uses a hypothesis test on the correlation and considers confounders such as stress or age.
Where marks are lost
E: using only r without testing or modelling. D: causal claims, self-report bias and the sample not being random. A: vague definition of the performance measure. C: a common topic with generic data.
Data
You need to collect enough honest responses and handle privacy and small expected counts.

The secretary problem and the 1/e stopping rule

Aim. For a sequence of n ranked candidates seen one at a time, how does the chance of picking the best or one of the top three change with the number rejected first, and how quickly does the best cut-off approach n/e?

  • AA and AI · SL and HL
  • Easy to start
  • Simulation
  • Common: on 2 sites

My take. Well known, so moderators will have seen it, but the derivation is solid. Make it yours with a real setting, such as choosing a flat from viewings, and with a variant of the success rule.

Mathematics, first steps and where marks are lost+
Mathematics
Conditional probability and the sum over positions of the best candidate; harmonic sums; comparing exact probabilities for n = 5 to 50 with integer optimisation; Riemann sum turning the sum into an integral of (1/x) to give x ln(1/x) and maximising by differentiation (HL); Monte Carlo simulation to check the exact results. Beyond syllabus: the dynamic programming view of the same rule.
How to start
Write out the exact success probability for n = 4 by listing all 24 orderings, then build a spreadsheet or Python function P(n,k) and simulate 100000 random permutations to confirm it. Then tabulate the best k for n up to 50.
SL and HL
SL students can do the exact discrete case, tabulate the best cut-off and compare with simulation. HL adds the integral limit, the derivation of 1/e by differentiation, and a variant where any of the top three counts as success, which changes the best rule.
Where marks are lost
E: stating the 37% rule without deriving the probability sum; B: unclear notation for 'best so far'; C: no personal context such as choosing a flat or a partner; D: not testing the rule's assumptions, such as no recall of rejected candidates.
Data
Pure theory plus a simple simulation; the only care needed is a clear derivation of the sum.

Are 30 percent rain forecasts really right 30 percent of the time?

Aim. How well calibrated are the daily rain probabilities issued for my town over a season, judged by observed outcomes?

  • AI · SL and HL
  • Needs care
  • Public data
  • Rarely listed

My take. A good question with real data and a clear answer. The twist is collecting the forecasts yourself, so start now.

Mathematics, first steps and where marks are lost+
Mathematics
Relative frequency, binned calibration table, conditional probability, binomial distribution, chi squared goodness of fit, Brier score as a mean of squared differences. Beyond syllabus: reliability diagram.
How to start
Record or download forecast rain probability and whether it rained for at least 100 days. Group forecasts into bins of 10 percent and compare each bin's stated probability with the observed frequency.
SL and HL
SL builds the calibration table, compares frequencies and computes the Brier score. HL adds a hypothesis test per bin with the binomial and a comparison of two forecasters with a test on the mean scores.
Where marks are lost
E: only counting days without any test. B: unclear definition of what counts as rain. C: data from nowhere in particular.
Data
Archived forecasts are harder to obtain than observations, so you may have to log them daily yourself.

Bayes and the prosecutor's fallacy in court

Aim. How far can a match in DNA evidence mislead a jury when the probability of a match given innocence is confused with the probability of innocence given a match?

  • AA and AI · SL and HL
  • Easy to start
  • Pure maths, no data
  • Rarely listed

My take. A sound choice if you vary the prior and treat several pieces of evidence. Keep the legal story short and let the algebra lead.

Mathematics, first steps and where marks are lost+
Mathematics
Conditional probability; tree diagrams; Bayes' theorem; sensitivity to the prior probability using tables and graphs of the posterior. Beyond syllabus: odds form of Bayes with likelihood ratios.
How to start
Choose one real case where a statistical argument was used in court, such as the Sally Clark case. Write the two conditional probabilities that were confused, then compute the posterior for a range of priors.
SL and HL
SL builds tree diagrams and computes posteriors for several priors. HL or top band work adds the odds form, combining several independent pieces of evidence, and how the posterior grows with each.
Where marks are lost
E: only one Bayes calculation, which is a textbook exercise. C: no reason why the case matters to the student. D: no discussion of how priors are chosen. A: a long case narrative that buries the maths.
Data
Pure theory with published figures; the danger is making up figures without justifying them.

Breaking Vigenere with chi-squared

Aim. How accurately can a chi-squared statistic against English letter frequencies recover the key length and key letters of a Vigenere cipher, and how does accuracy depend on the message length?

  • AA and AI · SL and HL
  • Easy to start
  • Simulation
  • Rarely listed

My take. Strong, because the statistics does real work and you can measure your own method's success. The simulation of accuracy against length gives it personal depth.

Mathematics, first steps and where marks are lost+
Mathematics
Modular arithmetic for shifts; relative frequencies; chi-squared statistic and its comparison with critical values; simulation of accuracy against text length. Beyond syllabus: index of coincidence as an alternative.
How to start
Take a public domain English text and encrypt extracts with random keys. Split a ciphertext into columns for a guessed key length and compute the chi-squared of each shift against English frequencies.
SL and HL
SL cracks a single cipher by hand and computes chi-squared for each shift. HL or top band work simulates thousands of ciphers, plots success rate against length and compares with the index of coincidence.
Where marks are lost
E: cracking one message shows only a procedure. B: code listings without explanation of the statistic. D: no discussion of short messages, where the test fails.
Data
All data can be generated by simulation; coding skill is the only requirement.

Building datasets with the same summary statistics but different shapes

Aim. Can I construct my own data sets that share mean, variance and correlation with a given set yet look completely different, and what does that say about summary statistics?

  • AA and AI · SL and HL
  • Easy to start
  • Pure maths, no data
  • Rarely listed

My take. Very familiar to moderators, so simply re-computing the quartet earns little. Constructing your own matching sets makes it personal and mathematical.

Mathematics, first steps and where marks are lost+
Mathematics
Mean, variance, Pearson correlation, least squares regression line, residuals, coefficient of determination. Beyond syllabus and reachable: solving simultaneous equations to force chosen statistics on a data set.
How to start
Compute the statistics for the four classic Anscombe sets. Then try to build a new set of 8 points with a curved pattern and adjust two points algebraically until the mean, variance and correlation match a target.
SL and HL
SL computes and compares the statistics and plots residuals. HL treats the construction as a system of equations, and can add rank correlation or a transformation that restores a linear fit.
Where marks are lost
C: only reproducing the famous example. E: stopping at calculating values already published. B: no explanation of why the correlation coefficient misses curvature.
Data
Small data sets by hand, so the only difficulty is the construction step, which needs careful algebra.

Do pea crosses follow Mendel's ratios? Testing with real counts

Aim. How well do published offspring counts from two-trait pea or fruit fly crosses agree with the 9:3:3:1 ratio predicted by independent assortment, and what does a linked-gene model change?

  • AI · SL and HL
  • Easy to start
  • Public data
  • Rarely listed

My take. It works when it tests a model against real counts, not as a probability exercise. The personal twist is picking a trait you can observe yourself, or comparing a case where the ratio holds with one where linkage breaks it.

Mathematics, first steps and where marks are lost+
Mathematics
Binomial and multinomial probability, conditional probability, tree diagrams and Punnett grids for dihybrid crosses, expected counts from a model, chi-squared test for goodness of fit with degrees of freedom and p-value (calculator), hypothesis testing. Beyond syllabus but reachable: a recombination fraction estimated by maximum likelihood, compared with the 9:3:3:1 model.
How to start
Download the count tables from Mendel's 1866 dihybrid peas (round/wrinkled, yellow/green: 315, 101, 108, 32) and a fruit fly linkage dataset such as Morgan's white eye and miniature wing crosses. First compute the probability of each phenotype from a Punnett grid, then scale to the sample size and compare with the observed numbers.
SL and HL
SL: build the probability model for one and two traits, then run a chi-squared test on the peas and interpret the p-value. HL depth: use the fly data where the model fails, fit a linkage parameter by maximising the likelihood, and discuss why Mendel's fit seems suspiciously good, with a clear look at the assumptions behind each test.
Where marks are lost
E: stopping at Punnett squares, which is basic probability, and never testing a model against data. B: not stating hypotheses, degrees of freedom or the significance level. C: a generic 'genetics is interesting' voice with no personal choice of organism or dataset. D: failing to question the independence assumption or the small sample effects.
Data
Classic count tables are publicly available and small, so there is no cleaning, but the student must check the expected counts are large enough for the test.

Do supermarket sugar bags really contain 1 kg?

Aim. Is the mean mass of 1 kg bags of sugar from different brands at least the labelled amount, and are the masses normally distributed?

  • AI · SL and HL
  • Needs care
  • Your own data
  • Rarely listed

My take. Good and practical if you compare brands and link the result to the legal rule; the personal touch could be a brand you buy regularly.

Mathematics, first steps and where marks are lost+
Mathematics
Mean and standard deviation, normal model, probability of a bag below 1000 g, two-sample t-test to compare brands, confidence interval for the mean and one-sample t-test against 1000 g (AI HL), a normal quantile plot or chi-squared goodness of fit. Beyond syllabus: the 'average quantity' packaging rule (the mean must not be below the label and few bags may fall below a tolerance).
How to start
Buy 30 or more bags of each of two brands and weigh each on a scale accurate to 1 g, checking its calibration with a known mass. Record the packaging mass separately if you weigh a bag opened.
SL and HL
SL models the masses as normal, estimates the proportion under 1000 g and compares two brands with a two-sample t-test. HL adds a confidence interval for each mean, a one-sample t-test against the label and a check of normality.
Where marks are lost
E: t-test used without checking the assumptions. A: unclear measuring method or calibration. D: no comment on scale error or whether the brand meets legal rules.
Data
Requires buying dozens of bags and a decent scale, but the data are quick to collect once done.

Do sweet colour proportions match the packet claims?

Aim. Do colour counts in several bags of sweets agree with the proportions stated by the manufacturer, and do people's favourite colours differ from what is in the bag?

  • AI · SL and HL
  • Easy to start
  • Your own data
  • Rarely listed

My take. Accessible but quite common and easily too shallow. The twist is the simulation of random bags to check whether the chi-squared results agree with what chance alone would give.

Mathematics, first steps and where marks are lost+
Mathematics
Chi-squared goodness-of-fit test with degrees of freedom and expected-count check, chi-squared test for independence on a survey table, hypotheses and p-value, combining bags into one sample. Beyond syllabus: multinomial confidence intervals or a simulation of random bags.
How to start
Buy at least 15 bags, count colours in each, and pool the totals. Separately, survey about 100 classmates for favourite colour. State the claimed proportions first.
SL and HL
SL applies the goodness-of-fit test to pooled counts and interprets the result. HL can add a test per bag, a simulation of the null model and a discussion of how bag-to-bag variation affects the test.
Where marks are lost
E: one test and nothing else. D: not noticing that comparing claims with preferences are two different questions. C: shallow motivation. A: unlabelled tables.
Data
Data is cheap to collect but the sample size must be large enough for expected counts of at least 5.

Do volcanic eruptions arrive at random? A Poisson check

Aim. Do eruptions of a chosen volcano, such as Etna or Kīlauea, occur as a Poisson process, judged by the counts per decade and the gaps between them?

  • AI · HL topic
  • Needs care
  • Public data
  • Rarely listed

My take. Good, provided you test the fit instead of assuming it. The twist is to compare a volcano that clusters with one that looks random.

Mathematics, first steps and where marks are lost+
Mathematics
Poisson distribution and its mean and variance, estimating λ from data, chi-squared goodness of fit with merged classes, comparison of variance with mean. Beyond syllabus: exponential waiting times between events.
How to start
Download the eruption record of one volcano from the Smithsonian Global Volcanism Program. Count eruptions starting in each decade, estimate λ, and compare the observed counts with Poisson predictions.
SL and HL
The Poisson distribution is AI HL. An SL student could use the binomial distribution on years with or without an eruption. HL depth comes from testing the fit and asking whether the variance exceeds the mean.
Where marks are lost
E: fitting λ and stopping there without a goodness of fit test. B: unclear definition of an eruption event. D: ignoring that older records are incomplete, which biases counts.
Data
Records are publicly available but need cleaning, and historical under-reporting must be handled explicitly.

Do wet days come in streaks? Testing rainfall models

Aim. Does a binomial or a Poisson model describe the number of wet days per month in one chosen city, and does a two-state Markov chain of dry and wet days explain the misfit better?

  • AI · SL and HL
  • Easy to start
  • Public data
  • Rarely listed

My take. Worth choosing because a model that fails for a physical reason makes a much better exploration than one that fits neatly. The personal twist is picking a place with real meaning for you, and comparing a wet-spell-prone climate with a steadier one to show the clustering is measurable.

Mathematics, first steps and where marks are lost+
Mathematics
Relative frequency estimates of p, binomial distribution B(n, p) with mean np and variance np(1-p), Poisson distribution with mean equal to variance, comparing observed counts with model counts using a chi-squared test for goodness of fit (degrees of freedom, merging small cells, p-value, significance level), variance-to-mean ratio as a quick diagnostic. Beyond syllabus if reachable: a two-state transition matrix for dry and wet days, with its long-run proportion found as an eigenvector (HL matrices, or simple simultaneous equations for SL), and the resulting run-length distribution being geometric.
How to start
Download daily precipitation records for one station over 25 to 30 years from a national meteorological archive, for example the Met Office historical station data, NOAA Climate Data Online or the ECA&D dataset. Choose a single month, such as October, to avoid seasonal mixing. Define a wet day with a stated threshold, say at least 1 mm. Count wet days per October for each year, compute the sample mean and variance, and estimate p as the total wet days divided by total days. Then tabulate the observed frequency of each count against the binomial prediction.
SL and HL
An SL student can do the binomial and Poisson fits, calculate the test statistic with the GDC or a spreadsheet, and interpret the result with a clear discussion of why independence fails. HL depth comes from building the Markov transition matrix from consecutive days, finding its steady state, simulating months from it, and showing the simulated counts fit the observed spread of counts far better than the binomial. Estimating the parameters from the same data and adjusting degrees of freedom shows top-band care.
Where marks are lost
A Presentation: lots of raw tables pasted in with no selection. B Mathematical communication: not defining a wet day, not stating hypotheses and degrees of freedom, unexplained GDC output. C Personal engagement: a generic city with no reason for the choice. D Reflection: stating the model fails without explaining why or testing the independence assumption. E Use of mathematics: a single test with expected counts below 5 left unmerged, or only a binomial fit with no second model.
Data
Daily rainfall archives are public and clean, though 25 or more years of data need a little tidying for missing days and a decision on the wet-day threshold.

Does caffeine or distraction change reaction time? A two sample test

Aim. Is the mean reaction time on an online ruler or screen test different between two conditions, once I allow for the spread in each group?

  • AI · SL and HL
  • Needs care
  • Your own data
  • Rarely listed

My take. A very common design, though it is workable if the condition is something you care about. Careful control of confounders is what separates a good one from a weak one.

Mathematics, first steps and where marks are lost+
Mathematics
Sample mean and standard deviation, two-sample t-test, hypotheses and p-value, checking normality with a histogram. Beyond syllabus for SL: the paired t-test (AI HL) and a permutation test.
How to start
Choose one condition, such as with and without music. Split about 30 volunteers at random between the two conditions, record the median of 10 trials for each person, plot the data, then run the test and check it against a GDC.
SL and HL
SL applies the two-sample t-test with clear hypotheses and interpretation. HL compares this with a paired design where everyone does both conditions, and with a permutation test, and discusses which design answers the question better.
Where marks are lost
E: running the test without checking assumptions or plotting. D: ignoring practice effects and small samples. A: cluttered raw data tables. C: no reason given for the chosen condition.
Data
Data must be collected with volunteers and consent, and practice effects need controlling.

Does retrieval practice beat rereading? A paired class experiment

Aim. Do students score higher on a short quiz after studying by self-testing than after rereading, and is the difference larger than chance variation?

  • AI · HL topic
  • Needs care
  • Your own data
  • Rarely listed

My take. Worth choosing when the design is paired and controlled. Comparing two different classes taught differently is a weak design.

Mathematics, first steps and where marks are lost+
Mathematics
Paired design, differences of scores, paired t-test with hypothesis and p-value, confidence interval for mean difference, checking normality with a histogram or normal plot. Beyond syllabus and reachable: effect size using Cohen's d.
How to start
Ask about 20 classmates for consent, give each two comparable word lists, one studied each way, and record scores. Get school approval first and anonymise names.
SL and HL
The paired t-test and confidence intervals are AI HL. HL work adds a confidence interval for the mean difference, an effect size and a discussion of order effects, with a check of the assumptions. An SL student could run a two-sample t-test, but that ignores the pairing, which should then be discussed honestly.
Where marks are lost
E: test run with no check of assumptions. A: messy tables. D: ignoring ethical and design limits. C: no personal link to studying.
Data
Collecting enough consented pairs takes effort and the sample stays small.

How many reviews before an average rating can be trusted

Aim. How many reviews does a product need before its mean rating settles within ±0.2 of the long-run value, and does this differ between two review sites?

  • AI · SL and HL
  • Needs care
  • Public data
  • Rarely listed

My take. Sound and doable, with a clear answer. The twist that makes it personal is choosing products or games you actually buy and checking whether the reviews would have changed your choice.

Mathematics, first steps and where marks are lost+
Mathematics
Sample mean and standard deviation, standard error, confidence intervals for a mean (t-interval), repeated random subsampling to show spread shrinking as 1/√n. HL adds a comparison of variances or a bootstrap-style argument and discussion of skewed, J-shaped rating distributions that break normality assumptions.
How to start
Pick 5 to 10 items with several hundred public reviews each, record the star ratings in a spreadsheet, then draw random subsamples of sizes 5, 10, 20, 50 and compute the spread of the means.
SL and HL
SL measures how the spread of subsample means shrinks with n and compares it with σ/√n. HL adds t-intervals for the mean (AI HL), tests the normal approximation on skewed data and compares two platforms.
Where marks are lost
E: stopping at descriptive means and one interval. B: not explaining why ratings are not normal. C: safe topic with little voice. D: ignoring selection bias in who writes reviews.
Data
Collecting a few hundred ratings by hand is slow, and platform terms may limit scraping, so use a dataset or manual sampling.

How many seats to oversell on a 180-seat flight

Aim. What number of tickets maximises expected profit for a 180-seat flight when passengers fail to show up independently with a stated probability, and how sensitive is this number to the costs?

  • AA and AI · SL and HL
  • Needs care
  • Public data
  • Rarely listed

My take. A clean optimisation with a real answer, and the sensitivity study is where the marks are. Weak if the costs are just invented.

Mathematics, first steps and where marks are lost+
Mathematics
Binomial distribution and its cumulative probabilities, expected value of a piecewise profit function, comparing consecutive expected values to find an integer optimum, sensitivity analysis. Beyond syllabus: a normal approximation to the binomial with a continuity correction, checked against the exact values.
How to start
Find a published no-show rate for airlines and set plausible ticket price and compensation cost. Compute expected profit for every ticket count from 180 to 200 in a spreadsheet using the binomial distribution.
SL and HL
SL: one no-show rate, one cost pair, exact binomial and a table. HL depth: normal approximation compared with exact values, sensitivity of the optimum to the no-show probability and cost ratio, and a test of the independence assumption using groups travelling together.
Where marks are lost
E: costs chosen with no source. B: profit function not defined precisely. D: independence assumption never questioned. C: no link to the student's own interests.
Data
Real no-show rates and compensation figures are patchy, so state clearly which values are assumed.

Is a basketball player's hot streak real?

Aim. Do a player's shooting results over a season show more streaks than a fixed shooting probability would produce by chance?

  • AI · SL and HL
  • Needs care
  • Public data
  • Rarely listed

My take. Turned into a streak test it has a clear answer and real statistics. Choose a team or player you follow.

Mathematics, first steps and where marks are lost+
Mathematics
Binomial distribution, probability of runs, simulation of the model, chi-squared test for independence between consecutive shots, expected value and variance. HL extension: Markov chain with make and miss states and its transition matrix. Beyond syllabus: the runs test.
How to start
Download shot-by-shot logs for one player from a public basketball statistics site for a full season. Compute overall shooting percentage, then the proportion of shots made after a make and after a miss.
SL and HL
SL compares the two conditional proportions with a chi-squared test and a simulation. HL builds a two-state Markov chain and compares its predictions with the data.
Where marks are lost
E: only means and percentages reported. B: hypothesis test steps not explained. D: no discussion of shot difficulty as a hidden variable. C: player chosen with no reason given.
Data
Shot logs are public but need cleaning into ordered sequences.

Is any lottery ever worth playing?

Aim. For which jackpot sizes, if any, does the expected return on a ticket in a national lottery exceed its price once rollovers and shared prizes are allowed for?

  • AA and AI · SL and HL
  • Easy to start
  • Public data
  • Rarely listed

My take. Common topic, so it only earns marks with the shared jackpot twist and a break-even analysis. Add historical draw data to check whether the number of winners fits your model.

Mathematics, first steps and where marks are lost+
Mathematics
Combinations to count outcomes for each prize tier; discrete random variable and expected value; binomial distribution for the number of jackpot winners; conditional expectation given a shared jackpot. Beyond syllabus for HL: Poisson approximation to the number of winners.
How to start
Take the official prize structure and ticket price of one lottery (for example EuroMillions or the UK National Lottery). Compute the probability of each tier with nCr, then the expected return for a fixed jackpot.
SL and HL
SL computes the expected value tier by tier and finds the break-even jackpot. HL adds the number of other winners as a binomial or Poisson variable, so the jackpot is split, and compares the corrected break-even figure with the naive one.
Where marks are lost
E: stopping at one expected value calculation with no modelling choice. B: unexplained notation for tiers. C: generic wording with no personal angle. D: no comment on what the model ignores, such as tax and utility.
Data
Prize tables and past jackpot sizes are published, so the data is easy; the sharing correction is the real work.

Is reaction time really normal? Testing a model on ruler-drop data

Aim. Do reaction times measured by an online test or ruler-drop follow a normal distribution, or is a skewed model such as the log-normal a better fit?

  • AA and AI · SL and HL
  • Needs care
  • Your own data
  • Rarely listed

My take. Fine, as long as the question is whether normality fails rather than an assumption that it holds. The skew is the interesting result.

Mathematics, first steps and where marks are lost+
Mathematics
Mean, standard deviation, histograms, normal probabilities and inverse normal, chi-squared goodness of fit test (AI), transformation of data by taking logarithms, comparison of quantiles. Beyond syllabus: the log-normal distribution.
How to start
Collect at least 100 reaction times from yourself and classmates using a ruler-drop or a simple browser test. Plot a histogram and compute the mean and standard deviation, then compare observed frequencies with those predicted by a fitted normal curve.
SL and HL
SL fits a normal model, checks the proportion within one and two standard deviations, and comments on skew. HL adds a formal goodness of fit test and a comparison with the log-normal fit.
Where marks are lost
E: assuming normality rather than testing it. A: too many raw tables. D: no discussion of measurement error, learning effects or outliers from lapses in attention.
Data
Needs enough clean repeated measurements, and the first few trials per person should be discarded because of practice effects.

Predicting league results with a Poisson goals model

Aim. How well does a Poisson model of goals, using team attack and defence strengths, predict the results of a full league season?

  • AI · HL topic
  • Easy to start
  • Public data
  • Rarely listed

My take. Popular, but a well-tested model with an honest goodness of fit check is respectable. Pick the student's own team and test the model on a season it was not fitted to.

Mathematics, first steps and where marks are lost+
Mathematics
Poisson distribution, expected value, independent events, chi-squared goodness of fit test, probability of a win, draw or loss by summing products. Fitting means from historical averages. HL extension: maximum likelihood ideas or a Markov chain for a league table.
How to start
Download one past Premier League season from football-data.co.uk, compute average goals for and against for each team, then compute expected goals for one match and the win, draw and loss probabilities.
SL and HL
The Poisson distribution is AI HL. HL students build the Poisson model, test it with chi-squared and compare predicted with real points; extra depth comes from the distribution of goal difference or from testing the model on a season it was not fitted to.
Where marks are lost
E: doing repeated arithmetic in a spreadsheet without justifying the Poisson assumption. D: failing to see that goals are not fully independent. C: generic, no personal team. A: pages of tables.
Data
The data is public, tidy and ready to use.

Screening tests and the base rate problem

Aim. How does the prevalence of a disease change the probability that a positive screening result is a true positive, and at what prevalence does a test become useful?

  • AA and AI · SL and HL
  • Needs care
  • Public data
  • Rarely listed

My take. A strong choice when the parameter study is the centre. Twist: use a test from a health issue you know, with sources cited.

Mathematics, first steps and where marks are lost+
Mathematics
Conditional probability, Bayes' theorem, tree diagrams, sensitivity and specificity, expected values for a screened population. Beyond syllabus: a comparison with a two-test sequence using independence assumptions.
How to start
Take published sensitivity, specificity and prevalence figures for a screening test from a public health agency such as the NHS or WHO, and compute the positive predictive value for a population of 10,000 people.
SL and HL
SL students can use tree diagrams and conditional probability with published figures. HL students can vary prevalence as a parameter, derive the PPV function, and analyse a two-stage screening design with its assumptions.
Where marks are lost
E: using only one calculation with no variation of parameters. D: no discussion of the independence assumption or data quality. A: long unstructured tables.
Data
Figures are public but need careful sourcing and a check that sensitivity and specificity come from the same study.

Size against sale price: linear, power or log model?

Aim. Which of a linear, power and logarithmic model best predicts the sale price of flats in my town from floor area, and how reliable is it outside the sampled range?

  • AI · SL and HL
  • Needs care
  • Public data
  • Rarely listed

My take. Common and easy to do badly. It becomes worthwhile through comparing several models, testing predictions on held-out sales and being honest about the many missing factors such as location.

Mathematics, first steps and where marks are lost+
Mathematics
Scatter plots, Pearson and Spearman correlation, least squares regression, log-log transformation for a power model and semi-log for a logarithmic one, residual plots, coefficient of determination, prediction error on held-out sales. Beyond syllabus: inference on the gradient and a second explanatory variable.
How to start
Download about 60 recent sales of one property type from a public sold-prices site or a listings portal in one postcode area. Plot price against area and check the residuals before choosing a model.
SL and HL
SL compares linear and power fits with residuals and interprets the coefficients. HL adds the logarithmic model through linearisation, tests predictions on held-out sales and discusses which model generalises best.
Where marks are lost
E: only a linear regression from a spreadsheet. C: little personal engagement when the topic is generic. D: no critique of outliers or sample choice. B: unexplained software output.
Data
Data are available but need cleaning, such as removing duplicates and mismatched property types.

Snakes and Ladders as a Markov chain: expected game length

Aim. For a real 100-square board, what is the expected number of turns to finish, and how does moving one ladder or snake change it?

  • AA and AI · SL and HL
  • Easy to start
  • Simulation
  • Rarely listed

My take. A good, achievable choice, but popular. Stand out by comparing several real commercial boards or by designing a board with a target game length.

Mathematics, first steps and where marks are lost+
Mathematics
Probability of dice outcomes, expected value, transition matrix, matrix powers and solving linear systems for the expected number of turns (AI HL matrices). Beyond syllabus: absorbing Markov chains with the fundamental matrix, and the geometric distribution as a simple model.
How to start
Code a 10-square toy board first and solve by hand. Then build the 100-square transition matrix and check it with a Monte Carlo simulation of 10000 games.
SL and HL
SL can use the toy board, tree diagrams and simulation. HL can handle the full matrix, solve for expected turns exactly and compare with simulation.
Where marks are lost
E: only simulation with no analytic result. B: unexplained code. D: no discussion of the exact-roll rule at the end. The 'strategy' claim is empty because players make no choices.
Data
Everything is computed from the board rules, though the finishing rule must be stated clearly.

Testing Zipf's law on English, Italian and Latin texts

Aim. How well does frequency proportional to 1/ranks fit word counts in three languages, and does the exponent s differ between them?

  • AI · SL and HL
  • Easy to start
  • Public data
  • Rarely listed

My take. A good choice with real data and a clear question. The personal twist is using a language you speak or your own writing, and being honest about where the law fails.

Mathematics, first steps and where marks are lost+
Mathematics
Log-log transformation, linear regression and least squares, correlation coefficient r, coefficient of determination, residual analysis, chi-squared goodness of fit. Beyond syllabus: comparison with a harmonic distribution.
How to start
Download three full novels from Project Gutenberg, count words with a short Python script or spreadsheet, rank them, and plot log frequency against log rank.
SL and HL
SL fits a straight line to the log-log data and interprets the gradient. HL examines the deviation at low and high ranks, tests the fit, and compares texts of different lengths.
Where marks are lost
E: regression done by software with no justification of the transformation. A: too many unlabelled plots. D: no discussion of why the tails deviate.
Data
Public domain texts are easy to count, but tokenisation choices such as apostrophes and accents must be stated.

Testing a claimed psychic with a binomial model

Aim. How many correct guesses out of n Zener-style card trials are needed before chance can be rejected at the 5% and 1% levels, and how does the power of that test change with the true success rate?

  • AA and AI · SL and HL
  • Easy to start
  • Your own data
  • Rarely listed

My take. Sound and accessible, but common as a hypothesis test, so the depth has to come from power and multiple testing. A twist: test yourself against a shuffled deck you control, then discuss why one lucky volunteer among many proves nothing.

Mathematics, first steps and where marks are lost+
Mathematics
Binomial distribution with P(X ≥ k) tail probabilities, null and alternative hypotheses, critical region, significance level, Type I and Type II errors. Beyond syllabus: power curve as a function of the true p, or a Bayesian comparison using a prior.
How to start
Design a card-guessing task with a fixed chance probability (for example five symbols, p = 0.2). Run 100 trials with 5 to 10 volunteers, then compute P(X ≥ observed) under the null hypothesis by hand and check with a calculator.
SL and HL
SL does the binomial tail probabilities and a one-tailed test with a clear critical region. HL or top-band work adds the power calculation for several alternative values of p and a discussion of multiple testing when many volunteers are tried.
Where marks are lost
E: stopping at 'p-value below 0.05 so psychic' without Type II error. D: no comment on the multiple comparison problem or on cheating and poor randomisation. C: generic if the student did not design their own trials.
Data
Data are easy to collect, though the number of trials must be planned in advance to avoid stopping when results look good.

Typical distance after n coin-flip steps

Aim. After n independent steps of ±1, how far from the start is a walker typically, and does the typical distance grow like √n?

  • AA and AI · SL and HL
  • Easy to start
  • Simulation
  • Rarely listed

My take. Strong if the exact calculation is done and compared with simulation. Twist: model a drunk walk on a real corridor with a wall, or a stock price.

Mathematics, first steps and where marks are lost+
Mathematics
Expectation and variance of sums of independent variables; binomial distribution for the position; E(X²)=n; probability of returning to the origin; simulation for verification. Beyond syllabus: Stirling's approximation for the central binomial coefficient (HL) and the normal approximation.
How to start
Write a short spreadsheet or Python simulation of 10 000 walks of n = 10, 100, 1000 steps, and record the mean absolute distance. Then compute E(X²) exactly and compare.
SL and HL
SL derives E(X)=0 and Var(X)=n and checks √n growth by simulation. HL uses the binomial distribution to compute E|X| exactly and applies Stirling to show it behaves like √(2n/π), or extends to biased steps.
Where marks are lost
E: relying on simulation only. B: confusing E|X| with √E(X²). A: unstructured code output. D: no comment on simulation error.
Data
Simulated data are easy to generate; the exact E|X| calculation needs care.

When to stop rolling: optimal strategy in a push-your-luck dice game

Aim. In a game where you keep adding dice totals until you roll a 1 and lose the turn, what stopping rule maximises the expected score per turn?

  • AA and AI · SL and HL
  • Easy to start
  • Simulation
  • Rarely listed

My take. A good accessible topic. Make it personal by using a game you actually play and by asking whether maximising expectation is the right aim when trailing.

Mathematics, first steps and where marks are lost+
Mathematics
Probability of independent events, expected value, geometric distribution, conditional probability, comparing the expected gain and loss of one more roll to find a threshold. Beyond syllabus: Markov chains for a whole game against an opponent.
How to start
Compute by hand the expected change in score from one more roll at a running total T (gain 4 x 5/6 on average, loss T x 1/6). Then simulate 100000 turns for each threshold to check it.
SL and HL
SL finds the threshold with expected values and a simulation. HL uses a recursion or a Markov chain, and can vary the die or the target score to model the winning probability rather than the score.
Where marks are lost
E: a single threshold found by trial and error with no algebra. B: unclear notation for states. D: no criticism of maximising the mean when a player only needs to win.
Data
Simulation is easy; exact results need care but stay tractable.

Where does e appear when random numbers are added up?

Aim. How many random numbers from 0 to 1 must be added on average for the total to exceed 1, and why is the answer e?

  • AA and AI · SL and HL
  • Easy to start
  • Simulation
  • Rarely listed

My take. Elegant and less overdone than most. Keep it anchored in the sum problem and make the proof yours.

Mathematics, first steps and where marks are lost+
Mathematics
Uniform continuous distribution, expected value, volume of the simplex 1/n! as a probability in n dimensions, E(N)=sum of P(N>n), Maclaurin series of ex, Monte Carlo simulation. Beyond syllabus: connection to the Irwin-Hall distribution.
How to start
Simulate 100000 trials in a spreadsheet or Python of adding uniform random numbers until the sum passes 1 and record the mean count. Then show P(sum of n numbers <1)=1/n! for n=2 and 3 by integration.
SL and HL
SL uses simulation plus n=2,3 areas and volumes and the known series for e. HL proves the general 1/n! by induction and integration and discusses convergence of the simulation.
Where marks are lost
E: presenting simulation as proof. B: unclear link between the hypercube and the sum. D: no error estimate for the simulated mean.
Data
Simulation is easy. The general proof by induction is the hard part but manageable.

Why doubling your bet never beats the house

Aim. Why does the doubling strategy give a near certain small win yet a negative expected gain, and how does a finite bankroll or table limit change the answer?

  • AA and AI · SL and HL
  • Easy to start
  • Simulation
  • Rarely listed

My take. A sound and pleasing choice if you show the exact calculation and the simulation agreeing. Keep away from claims about real trading, where data is hard to justify.

Mathematics, first steps and where marks are lost+
Mathematics
Geometric series and their sums, expectation of a discrete random variable, geometric distribution for the length of a losing run, probability of ruin, simulation of many sessions. Beyond syllabus: the martingale property and optional stopping.
How to start
Write down the profit after n consecutive losses, work out the probability of a run of length n on a fair and on a roulette style bet with p = 18/37, then simulate 10000 sessions with a bankroll cap.
SL and HL
SL computes the expected gain for a fixed number of rounds using geometric series and simulates. HL derives the expected gain analytically, explains why an unlimited bankroll seems to beat even a fair game by looking at the size of the stakes needed, and compares bankrolls with a table cap.
Where marks are lost
E: a simulation with no exact calculation to check it. B: unclear notation for a round and a run. D: not linking the finite result to the paradox. C: no distinct twist, such as applying it to a real trading style.
Data
Pure theory plus simple simulation; the maths is reachable and easy to control.

Ideas were collected from public lists and published exploration titles, merged when they are the same exploration, rewritten from scratch and checked by hand. “On N sites” counts how many public lists carry the same idea. How this list was made. Have a draft already? An IA review reads it against the five criteria before you submit.

Frequently asked questions

What are good IB Math IA ideas on statistics and probability?

+
Good starting points that are easy to begin and rarely listed are bayes and the prosecutor's fallacy in court, breaking Vigenere with chi-squared and building datasets with the same summary statistics but different shapes. Each has a clear question and room for a second method, which is where Personal engagement and Reflection come from.

Which statistics and probability IA topics are overdone?

+
monty Hall with n doors and a forgetful host, queueing at a real junction: arrivals, service and green time, can standing height predict a volleyball player's spike reach? and income and life expectancy across countries are either named in the subject reports among topics seen too often or appear on three or more public lists. They still work, but only with a twist that is your own.

Which of these statistics and probability ideas fit AA and which fit AI?

+
23 of the 42 ideas use mathematics from Analysis and Approaches and 42 from Applications and Interpretation; many fit both. 4 rely on HL content and are marked as HL.

Have a draft? Get it read before you submit.

I read your Internal Assessment in full, assess it criterion by criterion against the IB criteria, and discuss it with you in a 50-minute session. No AI, confidential. US$630.

How the IA review works