Free CAS MAS-I (Modern Actuarial Statistics I) Formula Sheet (2026)

Every MAS-I formula you need on the test, grouped by topic, rendered with full math notation. 103 formulas across 3 topics, calibrated to the 2026 syllabus. Free forever, no signup required.

103 Formulas
3 Topics
2026 Syllabus
Free Forever
Print-ready PDF: 1080x1350 portrait, math pre-rendered, fonts embedded. Download once, study anywhere.
Download PDF →

All MAS-I Formulas

Probability Models 31 items
Gamma waiting time density for the n-th arrival
fSn(s)=λnsn−1e−λs(n−1)!f_{S_n}(s) = \frac{\lambda^n s^{n-1} e^{-\lambda s}}{(n-1)!} — λ = rate, n = event index, s = waiting time, mean n/λ, variance n/λ²
Compound Poisson mean and variance
E[S(t)]=λt E[X], Var(S(t))=λt E[X2]E[S(t)] = \lambda t\,E[X],\ \text{Var}(S(t)) = \lambda t\,E[X^2] — λ = rate, t = time, X = iid severity, E[X²] = Var(X)+(E[X])²
NHPP mean function over an interval
Λ(a,b)=∫abλ(s) ds\Lambda(a,b) = \int_a^b \lambda(s)\,ds — λ(s) = time-varying intensity, [a,b] = interval; equals both mean and variance of N(b)−N(a)
Monte Carlo sample size for target half-width
n≈(1.96 s/h)2n \approx (1.96\,s/h)^2 — s = pilot sample SD, h = target half-width, n = required number of draws
Exponential inversion from a uniform
X=−1λln⁡(1−U)X = -\frac{1}{\lambda}\ln(1-U) — U = Uniform(0,1) draw, λ = exponential rate, X = exponential severity draw
Monte Carlo 95% confidence interval half-width
θ^n±1.96 s/n\hat\theta_n \pm 1.96\,s/\sqrt{n} — \hat\theta_n = sample mean of g(X_i), s = sample SD, n = independent draws
Inversion method draw from a uniform
X=F−1(U)X = F^{-1}(U) — U = Uniform(0,1) draw, F^{-1} = generalized inverse CDF, X = draw from target distribution F
Reversionary annuity to y after x dies
aˉy∣x=aˉy−aˉxy\bar a_{y|x} = \bar a_y - \bar a_{xy} — aˉy\bar a_y = single-life annuity on y, aˉxy\bar a_{xy} = joint-life annuity
Joint-life survival under common shock
tpxy=tpx∗⋅tpy∗⋅e−λt{}_tp_{xy} = {}_tp_x^* \cdot {}_tp_y^* \cdot e^{-\lambda t} — tp∗{}_tp^* = private survival, λ\lambda = shared hazard rate, t = time
Survival function from cumulative hazard
S(t)=exp⁡ ⁣(−∫0th(s) ds)=e−H(t)S(t) = \exp\!\left(-\int_0^t h(s)\,ds\right) = e^{-H(t)} — H(t) = cumulative hazard, h(s) = hazard rate, S(t) = survival probability
Last-survivor survival probability (inclusion-exclusion)
tpxy‾=tpx+tpy−tpxy{}_tp_{\overline{xy}} = {}_tp_x + {}_tp_y - {}_tp_{xy} — tpx{}_tp_x = prob x survives t, tpxy{}_tp_{xy} = joint survival
Mean residual life at age t
e(t)=∫t∞S(u) duS(t)e(t) = \frac{\int_t^\infty S(u)\,du}{S(t)} — S = survival function, t = current age, e(t) = expected remaining lifetime given survival to t
Joint-life continuous annuity under constant force and interest
aˉxy=1μx+μy+δ\bar a_{xy} = \dfrac{1}{\mu_x + \mu_y + \delta} — μx,μy\mu_x,\mu_y = constant forces of mortality, δ\delta = force of interest
Conditional survival probability (t-p-s)
tps=S(s+t)S(s){_t}p_s = \frac{S(s+t)}{S(s)} — S = survival function, s = current age, t = additional years survived
Hazard rate from density and survival
h(t)=f(t)S(t)=−ddtln⁡S(t)h(t) = \frac{f(t)}{S(t)} = -\frac{d}{dt}\ln S(t) — f = density, S = survival function, h = instantaneous failure rate
Constant-force whole life insurance EPV
Aˉx=μμ+δ\bar A_x = \frac{\mu}{\mu+\delta} — μ = constant force of mortality, δ = constant force of interest
Whole life insurance EPV (discrete)
Ax=∑k=0∞vk+1 k∣qxA_x = \sum_{k=0}^{\infty} v^{k+1}\,{}_{k|}q_x — v = 1/(1+i), k∣qx{}_{k|}q_x = prob of death in year k+1 for life age x
Insurance-annuity fundamental identity (discrete)
Ax=1−d a¨xA_x = 1 - d\,\ddot a_x — d = i/(1+i) effective discount rate, a¨x\ddot a_x = whole life annuity-due EPV, AxA_x = whole life insurance EPV
Whole life annuity-due EPV
a¨x=∑k=0∞vk kpx\ddot a_x = \sum_{k=0}^{\infty} v^{k}\,{}_k p_x — v = 1/(1+i), kpx{}_k p_x = prob life age x survives k years
Waiting time to the n-th event in a Poisson process
Sn=∑i=1nTi∼Gamma(n,λ)S_n = \sum_{i=1}^n T_i \sim \text{Gamma}(n,\lambda), E[Sn]=n/λE[S_n]=n/\lambda — T_i = iid Exponential(λ) gaps, n = event number, λ = rate
Non-homogeneous Poisson process mean function
N(b)−N(a)∼Poisson(∫abλ(s) ds)N(b)-N(a) \sim \text{Poisson}\left(\int_a^b \lambda(s)\,ds\right) — λ(s) = intensity function, [a,b] = time interval
Poisson process count probability
P(N(t)=k)=e−λt(λt)kk!P(N(t)=k) = \frac{e^{-\lambda t}(\lambda t)^{k}}{k!} — λ = rate, t = interval length, k = number of events
Compound Poisson aggregate loss mean and variance
E[S(t)]=λt E[X]E[S(t)] = \lambda t\, E[X], Var(S(t))=λt E[X2]\text{Var}(S(t)) = \lambda t\, E[X^2] — λ = rate, t = time, X = claim severity
Fundamental matrix of an absorbing Markov chain
N=(I−Q)−1N = (I - Q)^{-1} — I = identity, Q = transient-to-transient block of P, N_{ij} = expected visits to transient state j starting from i
Series system reliability with independent components
Rs=∏i=1nRiR_s = \prod_{i=1}^{n} R_i — R_i = reliability of component i, n = number of components in series
Bridge reliability by conditioning on the center element
Rbridge=RC⋅Rworks+(1−RC)⋅RfailsR_{\text{bridge}} = R_C \cdot R_{\text{works}} + (1 - R_C) \cdot R_{\text{fails}} — R_C = bridge element reliability, R_works = system reliability given C works, R_fails = given C fails
Parallel system reliability with independent components
Rp=1−∏i=1n(1−Ri)R_p = 1 - \prod_{i=1}^{n}(1 - R_i) — R_i = reliability of component i, 1 - R_i = unreliability, n = number of parallel components
Limited expected value (survival form)
E[X∧u]=∫0uS(x) dxE[X \wedge u] = \int_0^{u} S(x)\,dx — X = non-negative loss, u = cap/limit, S(x) = survival function 1−F(x)
Loss elimination ratio for ordinary deductible
LER(d)=E[X∧d]/E[X]\text{LER}(d) = E[X \wedge d] / E[X] — d = ordinary deductible, X = ground-up loss severity
Expected layer cost between deductible d and limit u
E[min⁡(X,u)−min⁡(X,d)]=E[X∧u]−E[X∧d]E[\min(X,u) - \min(X,d)] = E[X \wedge u] - E[X \wedge d] — X = loss, d = attachment, u = exhaustion point
Exponential limited expected value
E[X∧u]=θ(1−e−u/θ)E[X \wedge u] = \theta(1 - e^{-u/\theta}) — θ = exponential mean, u = policy limit/cap
Statistics 36 items
Collective risk model aggregate loss
S=∑i=1NXiS = \sum_{i=1}^{N} X_i — N = random claim count, X_i = iid severities independent of N
Compound distribution variance
Var(S)=E[N] Var(X)+Var(N) E[X]2\text{Var}(S) = E[N]\,\text{Var}(X) + \text{Var}(N)\,E[X]^2 — N = claim count, X = severity, S = aggregate loss
Panjer recursion for aggregate loss pmf
fS(sh)=11−afX(0)∑y=1s(a+b y/s) fX(yh) fS((s−y)h)f_S(sh) = \frac{1}{1 - a f_X(0)} \sum_{y=1}^{s} (a + b\,y/s)\, f_X(yh)\, f_S((s-y)h) — (a,b) = (a,b,0) class parameters, h = grid step
Normal approximation stop-loss premium
E[(S−d)+]=σ[ϕ(z)−z(1−Φ(z))]E[(S-d)_+] = \sigma[\phi(z) - z(1-\Phi(z))] — z = (d - E[S])/σ, σ = SD of S, φ = standard normal pdf, Φ = cdf
MLE of the exponential rate parameter
λ^=n/∑Xi=1/Xˉ\hat\lambda = n/\sum X_i = 1/\bar X — n = sample size, ΣX_i = sufficient statistic, X̄ = sample mean
Exponential family canonical density
f(x;θ)=h(x) c(θ)exp⁡(∑j=1kwj(θ) tj(x))f(x;\theta) = h(x)\,c(\theta)\exp\left(\sum_{j=1}^k w_j(\theta)\,t_j(x)\right) — h, c base functions; w natural params; t sufficient kernels
Fisher-Neyman factorization theorem
f(x1,…,xn;θ)=g(T(x),θ)⋅h(x)f(x_1,\ldots,x_n;\theta) = g(T(x),\theta)\cdot h(x) — T = sufficient statistic, g depends on θ through T, h depends only on data
UMVUE of theta for Uniform(0, theta)
θ^UMVUE=n+1nmax⁡Xi\hat\theta_{\text{UMVUE}} = \tfrac{n+1}{n}\max X_i — n = sample size, max X_i = largest order statistic (complete sufficient stat)
One-sample z test statistic for a mean
Z=Xˉ−μ0σ/nZ = \frac{\bar{X} - \mu_0}{\sigma/\sqrt{n}} — X-bar = sample mean, μ₀ = hypothesized mean, σ = population SD, n = sample size
Two-sided p-value for a z test
p=2⋅P(∣Z∣≥∣zobs∣)p = 2 \cdot P(|Z| \ge |z_{\text{obs}}|) — z_obs = observed test statistic, probability computed under H₀
Likelihood ratio test statistic
−2ln⁡(L0/L1)∼χk2-2\ln(L_0/L_1) \sim \chi^2_k — L₀ = restricted-model likelihood, L₁ = full-model likelihood, k = number of restrictions
Sample size for target power in a one-sided z test
n=σ2(z1−α+z1−β)2(μ0−μ1)2n = \frac{\sigma^2 (z_{1-\alpha} + z_{1-\beta})^2}{(\mu_0 - \mu_1)^2} — σ = SD, α = Type I rate, β = Type II rate, μ₀ = null mean, μ₁ = alternative mean
Likelihood contribution under left-truncation and right-censoring
Li(θ)=f(yi)δiS(yi)1−δiS(di)L_i(\theta) = \dfrac{f(y_i)^{\delta_i} S(y_i)^{1-\delta_i}}{S(d_i)} — f = density, S = survival, δᵢ = censoring indicator, dᵢ = left-truncation point
Nelson-Aalen cumulative hazard estimator
H^(t)=∑tj≤tsj/nj\hat H(t) = \sum_{t_j \le t} s_j/n_j — sⱼ = events at time tⱼ, nⱼ = risk-set size (counts only those with dᵢ < tⱼ ≤ yᵢ)
Conditional density under left-truncation at a deductible
fX∣X>d(x)=f(x)/S(d)f_{X\mid X>d}(x) = f(x)/S(d) for x>dx>d — f = unconditional density, S(d) = survival at deductible d
Likelihood for a right-censored sample
L(θ)=∏i=1nf(yi)δiS(yi)1−δiL(\theta) = \prod_{i=1}^{n} f(y_i)^{\delta_i} S(y_i)^{1-\delta_i} — f = density, S = survival, δᵢ = 1 if uncensored, 0 if right-censored at yᵢ
Likelihood contribution with left truncation and right censoring
Li(θ)=[f(xi∣θ)/S(di∣θ)]δi[S(ui∣θ)/S(di∣θ)]1−δiL_i(\theta) = [f(x_i\mid\theta)/S(d_i\mid\theta)]^{\delta_i}[S(u_i\mid\theta)/S(d_i\mid\theta)]^{1-\delta_i} — δ=1 if observed, d=truncation, u=censoring point
Fisher information for a single observation
I(θ)=E[(∂log⁡f/∂θ)2]=−E[∂2log⁡f/∂θ2]I(\theta) = E\left[\left(\partial \log f/\partial \theta\right)^2\right] = -E\left[\partial^2 \log f/\partial \theta^2\right] — f = density, θ = parameter
Cramer-Rao lower bound for an unbiased estimator
Var(θ^)≥1/[nI(θ)]\text{Var}(\hat\theta) \ge 1/[n I(\theta)] — n = sample size, I(θ) = Fisher information per observation, θ = parameter
Mean squared error decomposition
MSE(θ^)=Var(θ^)+[Bias(θ^)]2\text{MSE}(\hat\theta) = \text{Var}(\hat\theta) + [\text{Bias}(\hat\theta)]^2 — Bias(θ̂) = E[θ̂] − θ, Var = sampling variance of the estimator
CDF of the sample minimum
F(1)(x)=1−[1−F(x)]nF_{(1)}(x) = 1 - [1 - F(x)]^{n} — F = parent CDF, n = sample size, X_{(1)} = minimum
Unbiased sample variance with Bessel's correction
S2=1n−1∑i=1n(Xi−Xˉ)2S^2 = \frac{1}{n-1}\sum_{i=1}^{n}(X_i-\bar{X})^2 — n = sample size, X_i = i-th observation, Xˉ\bar{X} = sample mean, S² unbiased for σ²
Computational shortcut for sum of squared deviations
∑i=1n(Xi−Xˉ)2=∑i=1nXi2−nXˉ2\sum_{i=1}^{n}(X_i-\bar{X})^2 = \sum_{i=1}^{n} X_i^2 - n\bar{X}^2 — n = sample size, X_i = i-th observation, Xˉ\bar{X} = sample mean
Density of the k-th order statistic
f(k)(x)=n!(k−1)!(n−k)!F(x)k−1[1−F(x)]n−kf(x)f_{(k)}(x) = \frac{n!}{(k-1)!(n-k)!} F(x)^{k-1}[1-F(x)]^{n-k} f(x) — F = CDF, f = pdf, n = sample size, k = rank
CDF of the sample maximum
F(n)(x)=[F(x)]nF_{(n)}(x) = [F(x)]^{n} — F = parent CDF, n = sample size, X_{(n)} = maximum
Standard error of the sample mean
SE(Xˉ)=S/n\text{SE}(\bar{X}) = S/\sqrt{n} — S = sample standard deviation, n = sample size; uses σ/√n when σ known
Sample mean
Xˉ=1n∑i=1nXi\bar{X} = \frac{1}{n}\sum_{i=1}^{n} X_i — n = sample size, X_i = i-th observation, Xˉ\bar{X} = sample mean (unbiased for μ)
Uniform order statistic as a Beta distribution
U(k)∼Beta(k, n−k+1)U_{(k)} \sim \text{Beta}(k,\, n-k+1), E[U(k)]=k/(n+1)E[U_{(k)}] = k/(n+1) — n = sample size, k = rank, U = Uniform(0,1) sample
Z-test statistic for a single mean with known variance
Z=Xˉ−μ0σ/nZ = \dfrac{\bar{X} - \mu_0}{\sigma/\sqrt{n}} — Xˉ\bar{X} = sample mean, μ0\mu_0 = hypothesized mean, σ\sigma = known population SD, n = sample size
T-test statistic for a single mean with unknown variance
T=Xˉ−μ0s/n∼tn−1T = \dfrac{\bar{X} - \mu_0}{s/\sqrt{n}} \sim t_{n-1} — Xˉ\bar{X} = sample mean, μ0\mu_0 = hypothesized mean, s = sample SD, n = sample size
F-test statistic for the ratio of two variances
F=s12s22∼Fn1−1, n2−1F = \dfrac{s_1^2}{s_2^2} \sim F_{n_1-1,\,n_2-1} — s12,s22s_1^2, s_2^2 = sample variances (larger on top), n1,n2n_1, n_2 = sample sizes
Chi-square test statistic for a single variance
W=(n−1)s2σ02∼χn−12W = \dfrac{(n-1)s^2}{\sigma_0^2} \sim \chi^2_{n-1} — n = sample size, s2s^2 = sample variance, σ02\sigma_0^2 = hypothesized variance
Lognormal mean and variance
E[X]=eμ+σ2/2, Var(X)=e2μ+σ2(eσ2−1)E[X] = e^{\mu+\sigma^2/2},\ \text{Var}(X) = e^{2\mu+\sigma^2}(e^{\sigma^2}-1) — μ, σ = mean and SD of log X
Aggregate loss mean and variance (general two-term)
E[S]=E[N] E[X], Var(S)=E[N] Var(X)+Var(N) E[X]2E[S] = E[N]\,E[X],\ \text{Var}(S) = E[N]\,\text{Var}(X) + \text{Var}(N)\,E[X]^2 — N = claim count, X = iid severity
Negative binomial mean and variance
E[N]=rβ, Var(N)=rβ(1+β)E[N] = r\beta,\ \text{Var}(N) = r\beta(1+\beta) — r = shape, β = scale; variance exceeds mean by factor (1+β)
(a,b,0) class recursion
pk/pk−1=a+b/kp_k/p_{k-1} = a + b/k — p_k = probability of k claims, a and b = family-specific constants (Poisson, NegBin, Binomial)
Extended Linear Models 36 items
Pearson chi-square dispersion estimate
ϕ^=X2/(n−p)\hat\phi = X^2/(n-p) where X2=∑(riP)2X^2 = \sum (r_i^P)^2 — n = sample size, p = parameter count, r_i^P = Pearson residual
Deviance of a GLM
D=2[ℓ(y;y)−ℓ(μ^;y)]D = 2[\ell(\mathbf{y};\mathbf{y}) - \ell(\hat{\boldsymbol{\mu}};\mathbf{y})] — ℓ(y;y) = saturated log-likelihood, ℓ(μ̂;y) = fitted log-likelihood
Pearson residual for a GLM
riP=(yi−μ^i)/V(μ^i)r_i^{P} = (y_i - \hat\mu_i)/\sqrt{V(\hat\mu_i)} — y_i = observed, μ̂_i = fitted mean, V(μ̂_i) = variance function at μ̂_i
McFadden pseudo R-squared
RMcF2=1−ℓmodel/ℓnullR^2_{\text{McF}} = 1 - \ell_{\text{model}}/\ell_{\text{null}} — ℓ_model = fitted log-likelihood, ℓ_null = intercept-only log-likelihood
Incremental pure-premium GLM with base-rate offset
ln⁡E[Pi]=ln⁡(Bi)+β0+∑jβjxij\ln E[P_i] = \ln(B_i) + \beta_0 + \sum_j \beta_j x_{ij} — P = pure premium, B = current base premium, β = log-relativities to the base
Annualized Poisson claim frequency from an exposure-offset model
λ^i=μi/Ei=exp⁡(β0+∑jβjxij)\hat\lambda_i = \mu_i / E_i = \exp(\beta_0 + \sum_j \beta_j x_{ij}) — λ = per-exposure rate, μ = expected count, E = earned exposure
Population-averaged prediction across a control variable
μˉ=∑kpk g−1(ηk)\bar\mu = \sum_k p_k \, g^{-1}(\eta_k) — p_k = population share of control level k, η_k = linear predictor at level k, g = link function
Linear predictor in a GLM with an offset term
ηi=oi+β0+∑jβjxij\eta_i = o_i + \beta_0 + \sum_j \beta_j x_{ij} — o = known offset (coef fixed at 1), β = estimated coefficients, x = predictors, η = linear predictor
Score equation under the canonical link
X⊤(y−μ)=0X^\top (y - \mu) = 0 — X = design matrix, y = response vector, μ = fitted mean vector at the MLE
GLM response variance with dispersion and exposure weight
Var(Yi)=ϕ V(μi)/wi\mathrm{Var}(Y_i) = \phi\, V(\mu_i)/w_i — φ = dispersion, V = variance function, μ = mean, w = exposure weight
Log-link GLM with exposure offset
ln⁡μi=ln⁡(exposurei)+xi⊤β\ln \mu_i = \ln(\text{exposure}_i) + x_i^\top \beta — μ = mean response, exposure = policy-years at risk, x = covariates, β = coefficients
Exponential family density form
f(y;θ,ϕ)=exp⁡{(yθ−b(θ))/a(ϕ)+c(y,ϕ)}f(y;\theta,\phi) = \exp\{(y\theta - b(\theta))/a(\phi) + c(y,\phi)\} — θ = canonical parameter, φ = dispersion, b = cumulant function, a,c = known functions
Likelihood ratio statistic for nested GLMs
Λ=2(ℓ1−ℓ0)∼˙χΔp2\Lambda = 2(\ell_1 - \ell_0) \dot\sim \chi^2_{\Delta p} — ℓ₁ = full-model log-likelihood, ℓ₀ = reduced-model log-likelihood, Δp = extra parameters
Akaike information criterion for GLM selection
AIC=−2ℓ+2p\text{AIC} = -2\ell + 2p — ℓ = maximized log-likelihood, p = number of fitted parameters; lower is better
Elastic net penalized GLM objective
β^=arg⁡min⁡β{−ℓ(β)+λ[α∥β∥1+(1−α)∥β∥22]}\hat\beta = \arg\min_\beta \{-\ell(\beta) + \lambda[\alpha\|\beta\|_1 + (1-\alpha)\|\beta\|_2^2]\} — λ = penalty strength, α = L1/L2 mix (1 = lasso, 0 = ridge)
Bayesian information criterion for GLM selection
BIC=−2ℓ+pln⁡n\text{BIC} = -2\ell + p\ln n — ℓ = maximized log-likelihood, p = parameters, n = sample size; lower is better
Extended linear model linear predictor and link
η=β0+∑j=1pβjxj,  g(μ)=η\eta = \beta_0 + \sum_{j=1}^{p} \beta_j x_j,\; g(\mu)=\eta — η = linear predictor, β = coefficients, x = design columns, g = link, μ = mean response
Continuous-by-continuous interaction model
η=β0+β1x1+β2x2+β3(x1x2)\eta = \beta_0 + \beta_1 x_1 + \beta_2 x_2 + \beta_3 (x_1 x_2) — x1, x2 = continuous predictors, β3 = interaction coefficient measuring departure from additivity
Degrees of freedom for a categorical-by-categorical interaction
dfint=(k1−1)(k2−1)df_{int} = (k_1 - 1)(k_2 - 1) — k1, k2 = number of levels in the two categorical predictors; added on top of (k1-1)+(k2-1) main-effect df
Effective slope on x1 under a continuous interaction
∂η/∂x1=β1+β3x2\partial \eta / \partial x_1 = \beta_1 + \beta_3 x_2 — β1 = main-effect slope, β3 = interaction coefficient, x2 = partner predictor value
Standardized residual
ri=eis1−hiir_i = \dfrac{e_i}{s\sqrt{1-h_{ii}}} — e_i = raw residual, s = residual std error, h_ii = leverage of point i
Added variable plot residuals for predictor X_j
ey∣−j=y−y^(−j), ej∣−j=Xj−X^j,(−j)e_{y|-j} = y - \hat{y}_{(-j)},\ e_{j|-j} = X_j - \hat{X}_{j,(-j)} — hats with (-j) = fitted values from regressions that exclude X_j
Hat matrix for ordinary least squares
H=X(XTX)−1XTH = X(X^TX)^{-1}X^T — X = design matrix; h_ii = i-th diagonal of H is the leverage of observation i
Cook's distance for an observation
Di=ri2p+1⋅hii1−hiiD_i = \dfrac{r_i^2}{p+1} \cdot \dfrac{h_{ii}}{1-h_{ii}} — r_i = standardized residual, h_ii = leverage, p = number of predictors
Deviance-based pseudo R-squared for a GLM
Rdev2=1−DmodelDnullR^2_{\text{dev}} = 1 - \frac{D_{\text{model}}}{D_{\text{null}}} — D_model = deviance of fitted GLM, D_null = deviance of intercept-only model
Adjusted R-squared for linear regression
Radj2=1−(1−R2) n−1n−k−1R^2_{\text{adj}} = 1 - (1-R^2)\,\frac{n-1}{n-k-1} — n = sample size, k = number of slope parameters (intercept excluded)
Coefficient of determination for OLS with intercept
R2=1−SSESST=SSRSSTR^2 = 1 - \frac{\text{SSE}}{\text{SST}} = \frac{\text{SSR}}{\text{SST}} — SSE = error sum of squares, SST = total sum of squares, SSR = regression sum of squares
Scaled deviance of a GLM from log-likelihoods
D∗=2(ℓsat−ℓmodel)D^* = 2(\ell_{\text{sat}} - \ell_{\text{model}}), with unscaled D=ϕD∗D = \phi D^* — ℓ_sat = saturated log-likelihood, ℓ_model = fitted log-likelihood, φ = dispersion
Freedman-Diaconis rule for histogram bin width
h=2⋅IQR/n1/3h = 2 \cdot IQR / n^{1/3} — h = bin width, IQR = interquartile range of the data, n = sample size
Interquartile range
IQR=Q3−Q1IQR = Q_3 - Q_1 — Q1 = first quartile (25th percentile), Q3 = third quartile (75th percentile)
Tukey upper outlier fence for a box plot
Upper fence=Q3+1.5⋅IQR\text{Upper fence} = Q_3 + 1.5 \cdot IQR — Q3 = third quartile, IQR = interquartile range; points above are flagged outliers
Tukey lower outlier fence for a box plot
Lower fence=Q1−1.5⋅IQR\text{Lower fence} = Q_1 - 1.5 \cdot IQR — Q1 = first quartile, IQR = interquartile range; points below are flagged outliers
F statistic for analysis of deviance with estimated dispersion
F=(ΔD/Δdf)/ϕ^F = (\Delta D / \Delta df) / \hat\phi — ΔD = deviance reduction from added terms, Δdf = added parameters, φ̂ = estimated dispersion
Wald z statistic for a GLM coefficient
z=β^j/SE(β^j)z = \hat\beta_j / \text{SE}(\hat\beta_j) — β̂_j = MLE of coefficient j, SE = standard error of the estimate
GLM deviance
D=2[ℓ(saturated)−ℓ(β^)]D = 2[\ell(\text{saturated}) - \ell(\hat\beta)] — ℓ = log-likelihood, saturated = one parameter per observation, β̂ = fitted MLE
Pearson chi-square goodness-of-fit statistic for a GLM
X2=∑(yi−μ^i)2/V(μ^i)X^2 = \sum (y_i - \hat\mu_i)^2 / V(\hat\mu_i) — y_i = observation, μ̂_i = fitted mean, V = variance function

Frequently Asked Questions

Is the MAS-I formula sheet free?
Yes. The full MAS-I formula sheet is free, with no signup, no email, and no credit card required. 103 formulas across 3 topics, all rendered with the same KaTeX math notation used in the FreeFellow study app.
Can I download the MAS-I formula sheet as a printable PDF?
Yes. A 1080x1350 portrait PDF (Instagram and LinkedIn carousel native size, also great for tablet study) is linked at the top of this page. The PDF is fully self-contained: math is pre-rendered, fonts are embedded, no internet connection needed once downloaded.
What's covered on the MAS-I 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 CAS?
No. FreeFellow is not affiliated with the CAS or any examination body. This is an independent study aid covering the published syllabus.
What else is free at FreeFellow for MAS-I 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 MAS-I 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 an AI-native exam prep platform for actuarial (SOA & CAS), CFA, CFP, CPA, CAIA, and securities licensing candidates — built around modern AI as a core capability rather than as a bolt-on. Every lesson ships with AI-narrated audio. 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.