dvdbion.blogg.se

To make the models more robust, I've transformed the incoming past prices to relative changes, e.g.: would become, as 5250 = 5200 * (1+0.0096), and 5170 = 5250 * (1-0.0152). These models are also easy to implement even without using any external library after you get the trained weights. I wanted the model to use past data to try to predict probable future changes. I tried some MLP Neural Network models first, the Tanh layer also seemed ideal to output change percentages between. I knew that I could bring in some fancy stochastic model for these observations but I just went with the easier approach what I do in work anyways. I've made about 5 such observations so I would have enough data to test my model performances without waiting 6 seconds for every tick to happen. I've started with capturing the changes for a long period (200 ticks) for all the stocks. I've also created a generic algo trader that only needed a good predictive model so I could change approaches separately. I've captured "live" data to be able to replay them in these test environments. I've also created multiple testing environments to test my models. Even the forecast is changing stochastically.īased on the first impressions my general thought was that we can have an estimate of the underlying Outlook Magnitude based on past observations, as the changes are either seldom or slow enough so we can assume it to be constant given a limited a time-frame. These changes are gradually happening every tick, but these are also stochastic. Outlook Magnitude Forecast: defines the value Outlook Magnitude will tend to over time. These are steep changes in probability, but they are seldom, happening at about every 50-100 ticks. This unknown probability also changes over time, because of two factors:īull or Bear: changes the general direction of change (e.g.: 60% chance to increase would become 40% meaning it is more likely to decrease). Next it determines the volume of change which is again stochastic, based on the maximum volatility of the stock. The game first determines if it increases or decreases, which is based on some internal, unknown probability, called Outlook Magnitude.

I've dug through the game source code, most related files can be found here: SourceCode