Hi Craig,

I changed your difference relation slightly so y[n] is a function of in[[n]], otherwise the count on in and out does not match. The technique below uses the difference relation build a sum-of-squares error function between symbolic terms and the list of out values, and then uses FindMinimum to determine parameter values which minimize the error. It's a useful technique.

Best regards, David

In[1]:= in = {16.969, 15.151, 6.83824, 2.51202, 7.62716, 15.7341, 
   16.6111, 9.15233, 2.76235, 5.4981, 13.9102, 17.392, 11.5535, 
   3.75623, 3.83153, 11.6845, 17.4135, 13.7951, 5.39155, 2.79867, 
   9.28573, 16.6734, 15.6468, 7.50031, 2.50563, 6.96037, 15.2477, 
   16.9184, 9.86588, 2.98251};

In[2]:= out = {0.1, 0.235755, 0.520125, 1.58736, 5.30637, 10.8286, 
   14.5398, 17.9214, 24.4315, 36.8878, 47.4856, 51.8436, 54.3319, 
   57.7956, 63.7485, 68.5763, 70.3306, 71.0822, 71.8903, 73.2932, 
   74.6665, 75.2082, 75.4013, 75.5603, 75.8267, 76.1532, 76.3087, 
   76.3591, 76.3908, 76.4382};

In[3]:= Length /@ {in, out}

Out[3]= {30, 30}

In[4]:= (* mofified to give out[[n]] as a function of in[[n]]*)
y[n_] := y[n] = y[n - 1] (1 + a Exp[-(b in[[n]] + c (n))]);

In[5]:= y[0] = y0; (* to be determined *)

In[6]:= (* sum of squares error from measured out *)
(* to big to print conveniently *)
error = Sum[(y[n] - out[[n]])^2, {n, Length[out]}];

In[7]:= (* the first term of the sum *)
error[[1]]

Out[7]= (-0.1 + (1 + a E^(-16.969 b - c)) y0)^2

In[8]:= (* the second term *)
error[[2]]

Out[8]= (-0.235755 + (1 + a E^(-15.151 b - 2 c)) (1 + 
     a E^(-16.969 b - c)) y0)^2

In[9]:= (* minimize the error *)
sol = FindMinimum[error, {a, b, c, y0}]

Out[9]= {54.7861, {a -> 7.49923, b -> 0.0349481, c -> 0.296997, 
  y0 -> 0.0143244}}

In[10]:= (* a table of calculated out values *)
check = Table[y[n] /. sol[[2]], {n, Length[out]}];

In[11]:= plot = 
 ListPlot[{out, check}, PlotLegends -> {"given", "modeled"}]

Attachments:

I agree completely. The fit is reasonable, and it is easy to look at the error on a point-by-point basis. But there is no information regarding the fit parameters. There is no confidence interval associated with their values. In fact, one or more of the parameters could be insignificant. I usually think of curve fitting as a process for determining the values of parameters which are physically meaningful. It is the knowledge about the parameter values that is meaningful -- the graph is just a picture. So I think the deficiency is serious. I am spending some time looking at TimeSeriesModelFit in the documentation. I'm finding the water a bit murky. But it is definitely worth studying.

David, thank you very much for walking me through the FindMinimum approach to fitting parameters in my model. I played with your nb, and found that it does everything I wanted in terms of setting initial parameter estimates, gradient method, norm function ("error" in your nb). I'm looking at changing my modeling approach based on Sean's considered advice regarding well-known time series models. For now, though, you've provided the solution I was looking for. I appreciate your taking the time to help.

Sean, I saw your response after posting my method. It looks very interesting. I have not worked with TimeSeriesModelFit -- but I will certainly be investigating it. Thanks!