diff --git a/benchmarks/runbenchmarks.jl b/benchmarks/runbenchmarks.jl
index 9edf516..7284b30 100644
--- a/benchmarks/runbenchmarks.jl
+++ b/benchmarks/runbenchmarks.jl
@@ -23,7 +23,9 @@ for est_t in [MCMCEstimator, LSQEstimator, LinearApproxEstimator]
                 xs = rand($nsamplesdict[$size])
                 n = length(xs)
                 corrnoise = let n = 2*n
-                    1/10 * I(n) * MvNormal(zeros(n), I(n) + 1/5*hermitianpart(rand(n, n)))
+                    Σ = zeros(n, n)
+                    while !isposdef(Σ); Σ .= I(n) + 1/5*hermitianpart(rand(n, n)); end
+                    1/10 * I(n) * MvNormal(zeros(n), Σ)
                 end
                 noisemodel = CorrGaussianNoiseModel(corrnoise)
                 ysmeas = maybeflatten($f.(xs, [$θtrue])) .+ rand(corrnoise)
diff --git a/test/runtests.jl b/test/runtests.jl
index a26f180..4c4c503 100644
--- a/test/runtests.jl
+++ b/test/runtests.jl
@@ -92,8 +92,11 @@ end
         n = length(xs)
         # because observation are two dimensional
         corrnoise = let n = 2*n
-            1/10 * I(n) * MvNormal(zeros(n), I(n) + 1/5*hermitianpart(rand(n, n)))
+            Σ = zeros(n, n)
+            while !isposdef(Σ); Σ .= I(n) + 1/5*hermitianpart(rand(n, n)); end
+            1/10 * I(n) * MvNormal(zeros(n), Σ)
         end
+
         noisemodel = CorrGaussianNoiseModel(corrnoise)
 
         ysmeas = reduce(vcat, f.(xs, [θtrue])) .+ rand(corrnoise)