Two days are a small horizon. The random part is modelled by the binomial tree:
* [L] (1000-20)*r - 20 ... and finally ((1000-20)*r -20)*r
/
* [L] 1000-20 - * [K] (1000-20)*r ... and finally (1000-20)*r*r
/
1000 * * [L] 1000 * r - 20 ... and finally (1000*r -20)*r
\ /
* [K] 1000 - * [K] 1000 * r ... and finally 1000*r*r
with [L] used for LOSE, and [K] for KEEP.
Let $p$ be the probability for losing, i.e. for the [L] ramifications, and $q$ the complementary probability.
Above, the multiplication rate is (EDIT: adjusted after the edit in the OP)
$$
r = 1 + 5\% =1.05\ .
$$
The mean amount of money after two days is thus:
$$
\begin{aligned}
M
&=
p^2\cdot((1000-20)r-20)r
+pq\cdot(1000-20)r^2
\\
&\qquad\qquad
+pq\cdot(1000r-20)r
+q^2\cdot 1000r^2
\\
&=1000r^2-20p^2(r+r^2)-20pq(r+r^2)
\\
&=1000r^2-20p(r+r^2)(p+q)
\\
&=1000r^2-20p(r+r^2)
\ .
\end{aligned}
$$
Later EDIT:
An other way to think about this is as in the OP. The mean value after two days is by linearity $1000r^2$ plus mean value of the possible loss.
The probability to lose $20\$$ in the first day is $p$, and the corresponding mean is $p\cdot r^2\cdot 20\$$, using the convention that we "first lose, then use the daily rate".
The probability to lose $20\$$ in the second day is $p$ again, and the corresponding mean is $p\cdot r\cdot 20\$$, using the same convention.
So the mean, computed as in the OP comes to the same value:
$$1000 r^2 -20pr^2-20pr\qquad\text{(dollars)}\ .$$