Free SOA Exam PA (Predictive Analytics) Formula Sheet (2026)

Every Exam PA formula you need on the test, grouped by topic, rendered with full math notation. 53 formulas across 5 topics, calibrated to the 2026 syllabus. Free forever, no signup required.

53 Formulas
5 Topics
2026 Syllabus
Free Forever

All Exam PA Formulas

Predictive Analytics Problem Definition 7 items
Profit per policy
Profit=P−L−E\text{Profit} = P - L - E, P = premium (price) per policy, L = expected loss (claim severity times frequency), E = expense or acquisition cost per policy
Value as price minus cost
V=P−CV = P - C, V = value of the item, P = price (e.g. sale price per building), C = cost (e.g. construction cost per building)
Base rate of a binary target
p=neventNp = \dfrac{n_{event}}{N}, p = base rate (positive-class proportion), n_event = number of positive-class records, N = total records
Majority-class (no-information) accuracy
accmaj=1−pacc_{maj} = 1 - p, acc_maj = accuracy of always predicting the majority class, p = base rate of the rare (positive) class
Business impact of a predictive model
Impact=V×f\text{Impact} = V \times f, V = dollar value of the decision the prediction improves, f = how often that decision is made
Additive error data-generating model
y=f(x)+εy = f(x) + \varepsilon, y = observed response, f(x) = true signal function, ε\varepsilon = random noise with variance Var⁡(ε)\operatorname{Var}(\varepsilon)
Bias-variance decomposition of expected test MSE
E[(y0−f^(x0))2]=Var⁡(f^(x0))+[Bias⁡(f^(x0))]2+Var⁡(ε)E\left[(y_0 - \hat f(x_0))^2\right] = \operatorname{Var}(\hat f(x_0)) + \left[\operatorname{Bias}(\hat f(x_0))\right]^2 + \operatorname{Var}(\varepsilon), f^\hat f = fitted model, x0x_0 = test point, y0y_0 = true response, Var⁡(ε)\operatorname{Var}(\varepsilon) = irreducible error
Data Exploration and Visualization 8 items
Variance of a factor-level estimate
Var⁡(μ^level)∝1nlevel\operatorname{Var}(\hat\mu_{\text{level}}) \propto \frac{1}{n_{\text{level}}}, μ^level\hat\mu_{\text{level}} = estimated mean for that level, nleveln_{\text{level}} = number of rows in that level
Minority rows needed to oversample to a target share
f=pN1−pf = \frac{pN}{1-p} from p=fN+fp = \frac{f}{N+f}, p = target minority proportion, N = majority-class row count, f = minority rows after oversampling
Boxplot outlier fences
lower=Q1−1.5×IQR, upper=Q3+1.5×IQR\text{lower} = Q1 - 1.5\times\text{IQR},\ \text{upper} = Q3 + 1.5\times\text{IQR}, Q1 = first quartile, Q3 = third quartile, IQR = interquartile range
Interquartile range
IQR=Q3−Q1\text{IQR} = Q3 - Q1, Q1 = 25th percentile (first quartile), Q3 = 75th percentile (third quartile)
Shared variance between two predictors
r2=rxy2r^2 = r_{xy}^2, rxyr_{xy} = Pearson correlation between the two variables; r2r^2 = proportion of one variable's variation explained by the other
Sample variance
s2=1n−1∑i=1n(xi−xˉ)2s^{2} = \frac{1}{n-1}\sum_{i=1}^{n}(x_i - \bar{x})^{2}, xix_i = observation, xˉ\bar{x} = sample mean, n = sample size
Pearson correlation coefficient
rxy=∑(xi−xˉ)(yi−yˉ)∑(xi−xˉ)2∑(yi−yˉ)2r_{xy} = \frac{\sum (x_i - \bar{x})(y_i - \bar{y})}{\sqrt{\sum (x_i - \bar{x})^2}\sqrt{\sum (y_i - \bar{y})^2}}, x,y = paired values, xˉ,yˉ\bar{x},\bar{y} = means, i = observation index
Sample standard deviation
s=s2=1n−1∑i=1n(xi−xˉ)2s = \sqrt{s^{2}} = \sqrt{\frac{1}{n-1}\sum_{i=1}^{n}(x_i - \bar{x})^{2}}, s2s^2 = sample variance, xix_i = observation, xˉ\bar{x} = mean, n = sample size
Data Transformations and Unsupervised Learning Techniques 12 items
Standardized feature (z-score for clustering)
z=x−μσz = \dfrac{x - \mu}{\sigma}, x = raw feature value, μ = feature mean, σ = feature standard deviation
Complete linkage between-cluster distance
d(A,B)=max⁡a∈A, b∈B∥a−b∥d(A,B) = \max_{a \in A,\, b \in B} \lVert a - b \rVert, A, B = clusters, a, b = member points, distance = farthest pair
Proportion of variance explained by a principal component
PVEm=Var(zm)∑jVar(zj)\text{PVE}_m = \frac{\text{Var}(z_m)}{\sum_j \text{Var}(z_j)}, PVEm = proportion for PC m, Var(zm) = variance of PC m, denominator = total variance across all PCs
Variance of a principal component from its standard deviation
Var(zm)=sm2\text{Var}(z_m) = s_m^2, Var(zm) = variance captured by PC m, sm = reported standard deviation of PC m; total variance = number of standardized variables
Within-cluster sum of squares (K-means objective)
WCSS=∑k=1K∑x∈Ck∥x−μk∥2\text{WCSS} = \sum_{k=1}^{K} \sum_{x \in C_k} \lVert x - \mu_k \rVert^{2}, K = number of clusters, C_k = cluster k, x = observation, μ_k = mean vector of cluster k
Single linkage between-cluster distance
d(A,B)=min⁡a∈A, b∈B∥a−b∥d(A,B) = \min_{a \in A,\, b \in B} \lVert a - b \rVert, A, B = clusters, a, b = member points, distance = closest pair
First principal component score
z1=ϕ11x1+ϕ21x2+ϕ31x3z_1 = \phi_{11}x_1 + \phi_{21}x_2 + \phi_{31}x_3, z1 = PC1 score, φj1 = loading of variable j on PC1, xj = standardized feature j
Loading vector normalization constraint
∑jϕjm2=1\sum_{j} \phi_{jm}^2 = 1, φjm = loading of variable j on PC m; squared loadings of a component sum to one (unit-length loading vector)
Number of dummy columns for a categorical variable
d=k−1d = k - 1, k = number of category levels, d = number of dummy (indicator) columns needed to avoid redundancy in a GLM
Log transform for a skewed non-negative variable
x′=ln⁡(x+1)x' = \ln(x + 1), x = raw strictly non-negative value, x' = transformed value; the +1 lets x = 0 be handled
Min-max scaling
xmm=x−xmin⁡xmax⁡−xmin⁡x_{\text{mm}} = \dfrac{x - x_{\min}}{x_{\max} - x_{\min}}, x = raw value, xmin⁡x_{\min} = minimum, xmax⁡x_{\max} = maximum, xmmx_{\text{mm}} = scaled value in [0,1]
Z-score standardization
z=x−xˉsz = \dfrac{x - \bar{x}}{s}, x = raw value, xˉ\bar{x} = sample mean, s = sample standard deviation, z = standardized value
Generalized Linear Models 15 items
Poisson log-link model with a log-exposure offset
log⁡(E[count])=log⁡(exposure)+β0+β1x1+⋯\log(E[\text{count}]) = \log(\text{exposure}) + \beta_0 + \beta_1 x_1 + \cdots, exposure = units of exposure (offset, coef fixed at 1), β0\beta_0 = intercept, βj\beta_j = slope, xjx_j = predictor
Weighted least squares objective
min⁡β∑iwi(yi−y^i)2\min_{\beta} \sum_{i} w_i (y_i - \hat{y}_i)^2, wiw_i = weight of row i, yiy_i = observed response, y^i\hat{y}_i = fitted value, β\beta = coefficients
Group-size weighted mean of grouped rates
yˉw=∑iniyi∑ini\bar{y}_w = \dfrac{\sum_i n_i y_i}{\sum_i n_i}, nin_i = group size (weight) of row i, yiy_i = group-average rate for row i
Expected count as exposure times a fitted rate
E[count]=exposure×eβ0+β1x1E[\text{count}] = \text{exposure} \times e^{\beta_0 + \beta_1 x_1}, exposure = units of exposure, eβ0+β1x1e^{\beta_0 + \beta_1 x_1} = rate per unit exposure, β0\beta_0 = intercept, β1\beta_1 = slope, x1x_1 = predictor
GLM log link function
log⁡(μ)=β0+β1x1+…\log(\mu) = \beta_0 + \beta_1 x_1 + \dots, so eβ1e^{\beta_1} is the factor by which a one-unit rise in x1x_1 scales μ\mu; μ\mu = expected response, β\beta = coefficients
Pearson residual for a GLM
ri=yi−μ^iV(μ^i)r_i = \frac{y_i - \hat{\mu}_i}{\sqrt{V(\hat{\mu}_i)}}, yiy_i = observed, μ^i\hat{\mu}_i = fitted mean, V(μ^i)V(\hat{\mu}_i) = variance function at the fitted mean
Lognormal mean back-transformation (smearing correction)
E^[Y]=eμ^log⁡ eσ2/2\hat{E}[Y] = e^{\hat{\mu}_{\log}}\, e^{\sigma^2/2}, μ^log⁡\hat{\mu}_{\log} = fitted value on log scale, σ2\sigma^2 = residual variance on log scale
Estimated dispersion parameter
ϕ^=1n−p∑i=1n(yi−μ^i)2V(μ^i)\hat{\phi} = \frac{1}{n-p}\sum_{i=1}^{n}\frac{(y_i - \hat{\mu}_i)^2}{V(\hat{\mu}_i)}, n = observations, p = parameters, yiy_i = observed, μ^i\hat{\mu}_i = fitted mean, VV = variance function
Elastic net penalized regression objective
min⁡β12n∑i(yi−β0−∑jβjxij)2+λ[1−α2∑jβj2+α∑j∣βj∣]\min_{\beta} \frac{1}{2n}\sum_i (y_i - \beta_0 - \sum_j \beta_j x_{ij})^2 + \lambda[\frac{1-\alpha}{2}\sum_j \beta_j^2 + \alpha\sum_j |\beta_j|], λ = penalty strength, α = L1/L2 mix, β = coefficients
Variance-inflation factor for a predictor
VIFj=11−Rj2\text{VIF}_j = \frac{1}{1 - R_j^2}, R_j^2 = R-squared from regressing predictor j on all other predictors
One-standard-error threshold for lambda.1se selection
threshold=MSEmin⁡+SEmin⁡\text{threshold} = \text{MSE}_{\min} + \text{SE}_{\min}, MSE_min = minimum CV mean squared error, SE_min = its standard error; lambda.1se = largest λ with CV MSE ≤ threshold
Log-link multiplicative effect of a coefficient
μnew=μ×eβ\mu_{new} = \mu \times e^{\beta} per unit, μ = expected response, β = coefficient on the link scale; eβ−1e^{\beta}-1 is the percent change
Logit-link odds ratio from a coefficient
OR=eβOR = e^{\beta}, OR = multiplicative change in the odds of the event per unit increase, β = logit-link coefficient
Interaction-adjusted slope of a predictor at a level
slope of x1 at level ℓ=β1+βint,ℓ\text{slope of } x_1 \text{ at level } \ell = \beta_1 + \beta_{\text{int},\ell}, β1 = main coefficient, β_int,ℓ = interaction coefficient (zero for baseline level)
GLM linear predictor with link function
g(μ)=η=β0+β1x1+…g(\mu) = \eta = \beta_0 + \beta_1 x_1 + \dots, g = link function, μ = expected response, η = linear predictor, β = coefficients, x = predictors
Tree-Based Models 11 items
Default mtry for random forest classification
mtry=pm_{try} = \sqrt{p}, p = total number of predictors, m_try = predictors sampled at each split
Out-of-bag probability a row is left out of a bootstrap sample
P=(1−1/n)n→e−1≈0.368P = (1-1/n)^{n} \to e^{-1} \approx 0.368, n = number of training rows, P = chance a given row is omitted from one bootstrap (~37%)
Default mtry for random forest regression
mtry=p/3m_{try} = p/3, p = total number of predictors, m_try = predictors sampled at each split
Gradient boosting additive update
Fm(x)=Fm−1(x)+λ hm(x)F_m(x) = F_{m-1}(x) + \lambda\, h_m(x), F = ensemble prediction, m = boosting round, λ = learning rate (shrinkage), h_m = tree fit to current pseudo-residuals
k-fold cross-validation error
CV error=1k∑i=1kError(foldi)\text{CV error} = \frac{1}{k}\sum_{i=1}^{k} \text{Error}(\text{fold}_i), k = number of folds, fold_i = i-th held-out validation fold
Variance of an averaged tree ensemble
Var⁡=ρσ2+1−ρBσ2\operatorname{Var} = \rho\sigma^{2} + \frac{1-\rho}{B}\sigma^{2}, ρ = pairwise tree correlation, σ² = single-tree prediction variance, B = number of trees
Learning-rate to tree-count inverse tradeoff
nnew≈nold×λoldλnewn_{new} \approx n_{old} \times \frac{\lambda_{old}}{\lambda_{new}}, n = number of trees, λ = learning rate; halving λ roughly doubles the trees needed
Cost-complexity pruning criterion
Rcp(T)=R(T)+cp⋅∣T∣R_{cp}(T) = R(T) + cp \cdot |T|, R(T) = tree total error, cp = complexity parameter, |T| = number of terminal nodes (leaves)
Gini impurity of a node
G=1−∑kpk2G = 1 - \sum_{k} p_k^{2}, G = Gini impurity, p_k = proportion of class k in the node; G = 0 at a pure node (rpart default)
Entropy of a node
H=−∑kpklog⁡2pkH = -\sum_{k} p_k \log_2 p_k, H = entropy in bits, p_k = proportion of class k in the node; H = 0 at a pure node
Root mean square error
RMSE=1n∑i(yi−y^i)2\text{RMSE} = \sqrt{\frac{1}{n}\sum_i (y_i - \hat{y}_i)^2}, y_i = actual value, y^i\hat{y}_i = predicted value, n = number of observations

Frequently Asked Questions

Is the Exam PA formula sheet free?
Yes. The full Exam PA formula sheet is free, with no signup, no email, and no credit card required. 53 formulas across 5 topics, all rendered with the same KaTeX math notation used in the FreeFellow study app.
Will there be a printable PDF version?
A printable PDF is rolling out shortly. In the meantime, the inline page below is print-friendly: most browsers print clean copies via the Print menu (the navigation, footer, and download CTA are hidden in print).
What's covered on the Exam PA formula sheet?
Every formula is grouped by official syllabus topic, with the formula in math notation plus a one-line note on when to use it (or a watch-out from CAIA, CFA, or other prep-provider commentary). Coverage is calibrated to the 2026 syllabus and refreshed when the corpus changes.
What is FreeFellow's relationship with SOA?
No. FreeFellow is not affiliated with the SOA or any examination body. This is an independent study aid covering the published syllabus.
What else is free at FreeFellow for Exam PA candidates?
The full question bank with detailed solutions, mixed practice, readiness tracking, lessons (where available), and the formula sheet are all free forever. Fellow ($59/quarter or $149/year per track) unlocks timed mock exams, spaced-repetition flashcards, performance analytics, AI essay grading, and a personalized study plan.
Practice Exam PA questions free →

Formula sheets for other exams

Exam P Exam FM Exam FAM Exam SRM Exam ALTAM Exam ASTAM CFA Level I CFA Level II CFA L3 Portfolio Mgmt CFA L3 Private Wealth CFA L3 Private Markets CFP

About FreeFellow

FreeFellow is a free exam prep library for actuarial (SOA & CAS), CFA, CFP, CPA, CAIA, GARP FRM, IRS Enrolled Agent, IMA CMA, and FINRA / NASAA securities licensing candidates. The entire question bank, written solutions, and lessons are free for every candidate, with no trial period and no credit card. Lessons include narrated audio, and every constructed-response item has a copy-to-AI prompt builder so candidates can paste their answer into their own ChatGPT or Claude for self-graded feedback; Fellow members get instant AI grading on essays against the official rubric (currently CFA Level III, expanding to other essay-bearing sections).

The 70% you need to pass (question bank, written solutions, lessons, formula sheet, mixed practice, readiness tracking) is free forever, with no trial period and no credit card. Become a Fellow ($59/quarter or $149/year per track) to unlock mock exams, flashcards with spaced repetition, performance analytics, AI essay grading, and a personalized study plan.