The power of neural networks is that they can learn, given the desired output of one set of inputs, how to find the desired output to other sets of inputs. Backpropogation is just one stage in that learning process, but the only one it has in common with markets.
Markets have a relatively fixed set of inputs, and use a backpropogation like mechanism to find the optimal weights for those inputs. But that doesn’t require a neural network to do—it’s super trivial maths. I haven’t seen any evidence that the neural network like structure of a market let’s it guess the optimal quantity of X to produce given an entirely new set of inputs without requiring a feedback mechanism.
There are definitely differences. One is that NNs are trained on training data and then let loose on real world (or testing) data. Markets are always training online. Another is that NNs (are supposed to) approximate a true hidden function, whereas markets are adapting to changing conditions not necessarily to a single underlying truth. But markets do adapt to inputs they haven’t seen before, and there are economic theories describing that process, like adaptive expectations and tatonnement. I suspect that markets are more likely to adjust quite quickly, and also to “forget” old data quite quickly.
I don’t think this analogy holds at all.
The power of neural networks is that they can learn, given the desired output of one set of inputs, how to find the desired output to other sets of inputs. Backpropogation is just one stage in that learning process, but the only one it has in common with markets.
Markets have a relatively fixed set of inputs, and use a backpropogation like mechanism to find the optimal weights for those inputs. But that doesn’t require a neural network to do—it’s super trivial maths. I haven’t seen any evidence that the neural network like structure of a market let’s it guess the optimal quantity of X to produce given an entirely new set of inputs without requiring a feedback mechanism.
There are definitely differences. One is that NNs are trained on training data and then let loose on real world (or testing) data. Markets are always training online. Another is that NNs (are supposed to) approximate a true hidden function, whereas markets are adapting to changing conditions not necessarily to a single underlying truth. But markets do adapt to inputs they haven’t seen before, and there are economic theories describing that process, like adaptive expectations and tatonnement. I suspect that markets are more likely to adjust quite quickly, and also to “forget” old data quite quickly.