diff --git a/poincare/src/integer.cpp b/poincare/src/integer.cpp
index 575fb1357..b712e4f2d 100644
--- a/poincare/src/integer.cpp
+++ b/poincare/src/integer.cpp
@@ -308,6 +308,15 @@ float Integer::approximate(Context& context) const {
     mantissa |= (beforeLastDigit >> numberOfBitsInLastDigit);
   }
 
+  if ((m_numberOfDigits==1) && (m_digits[0]==0)) {
+    /* This special case for 0 is needed, because the current algorithm assumes
+     * that the big integer is non zero, thus puts the exponent to 126 (integer
+     * area), the issue is that when the mantissa is 0, a "shadow bit" is
+     * assumed to be there, thus 126 0x000000 is equal to 0.5 and not zero.
+     */
+    return m_negative ? -0.0f : 0.0f;
+  }
+
   uint_result = 0;
   uint_result |= (sign << 31);
   uint_result |= (exponent << 23);
diff --git a/poincare/test/integer.cpp b/poincare/test/integer.cpp
index 64d9de213..ac704b0fa 100644
--- a/poincare/test/integer.cpp
+++ b/poincare/test/integer.cpp
@@ -54,4 +54,7 @@ QUIZ_CASE(poincare_integer_approximate) {
   Context context;
   assert(Integer(1).approximate(context) == 1.0f);
   assert(Integer("12345678").approximate(context) == 12345678.0f);
+  assert(Integer("0").approximate(context) == 0);
+  assert(Integer("-0").approximate(context) == -0);
+  assert(Integer(-1).approximate(context) == -1);
 }