Search for optimal alpha in EWMA All literature about finding the best alpha for a EWMA points to use RMSE to measure the fit between the EWMA and the signal. As alpha increases, the series get less and less smoothed out, and as a consequence, the smoothed values get more and more similar to the signal, leading to an optimization that always select the highest alpha possible.
An alternative to fitting the entire signal with the smoothed series is to do an one-step-ahead forecast, and compare the signal value with the predicted one, but still it seems that this will optimize for the highest alpha possible. Is there something that I´m missing here?
EDIT adding some code to exemplify what I´m talking about. First, let´s create an fake signal:
idx = pd.date_range('2016-01-01', '2016-05-04', freq='D')
phase = np.pi / 2
len_idx = len(idx)
points = np.arange(0, len_idx)

and let´s calculate the EMWA:
def ewm(y, alpha):
    cur_ewm = np.zeros([len(y)])
    cur_ewm[0] = y[0]
    for i in range(1, len(y)):
        cur_ewm[i] = alpha * y[i] + (1-alpha)*cur_ewm[i-1]
    return cur_ewm

df['y'] = np.sin(points * 2 * np.pi / 30 + phase) + points / 100 + np.random.normal(size=len_idx)
df.index = idx

mean_squared_error(df.y, ewm(df.y, 0.8))
0.05528982348082824

mean_squared_error(df.y, ewm(df.y, 0.6))
0.19605838797331043

mean_squared_error(df.y, ewm(df.y, 0.1))
1.039271870184871

as the alpha goes up, the mse goes down. 
 A: Although the forecast and actual series become more "similar" in terms of appearance to the human eye as $\alpha \rightarrow 1$, they don't, in general, become more "similar" in terms of "smaller RMSE".  You can use RMSE, MAE, MAPE, or a wide range of other criteria and you will still not see the effect you describe.  The smoothed values do not automatically get more and more similar to the signal at each individual period as $\alpha \rightarrow 1$, they get more and more similar to the previous period's signal.  RMSE and other metrics measure the similarity at each individual period, not whether a series, if shifted by some number of periods, is the same as another series. The overall pattern will appear the same, but shifted by one period - and that shift means that you can do better with lower $\alpha$ values, e.g., 0.1 - 0.3, in the bulk of cases.
We'll construct a series of 100 observations that are i.i.d. Gaussian and plot a curve relating RMSE and $\alpha$:
x <- rnorm(110)
alpha <- seq(0.05,1,by=0.05)
rmse <- rep(0,length(alpha))

for (i in 1:length(alpha)) {
   # Initialization 
   f <- rnorm(1)
   for (j in 1:9) {
      f <- (1-alpha[i])*f + alpha[i]*x[j]
   }
   # Evaluation
   for (j in 10:99) {
      f <- (1-alpha[i])*f + alpha[i]*x[j]
      e <- x[j+1] - f
      rmse[i] <- rmse[i] + e*e
   }
}
rmse <- sqrt(rmse/100)
plot(rmse~alpha, type='b', lwd=2, pch=16)


As our series gets longer, our optimal $\alpha$ will, in this case, get smaller.  Other series will, of course, have other optimal values for $\alpha$, but it is not, in general, the case that $\alpha = 1$ is optimal.
