1

Topic: Forecasting of sequence of events

Colleagues, It is necessary to predict sequence of events in time. We assume to eat events ' A ', ' B ', ' a C ' etc. the Latin alphabet. These events happen in time so that chain ABCDABDSSABCSSABCSS To me this chain is formed it is necessary to continue. As you can see in an example of all some events A, B, a C, D, SS. And it is easy to note that ABC, AB - follow one after another. I.e. if  event A it is possible to state with high probability approach BC and so on. Events Z, F, K - do not meet at all. For each event there is an approach date. How such tasks dare? Circuits ?  neural networks?  on answers for the task above - I will be already grateful. But actually the task is more difficult. To each event at its approach the number from 1 to 0 - fractional is compared. And I need to predict this number.

2

Re: Forecasting of sequence of events

Hello, Gattaka, you wrote: G> Colleagues, G> It is necessary to predict sequence of events in time. We assume to eat events ' A ', ' B ', ' a C ' etc. the Latin alphabet. G> these events happen in time so that chain G> G> ABCDABDSSABCSSABCSS G> To me this chain is formed it is necessary to continue. As you can see in an example of all some events A, B, a C, D, SS. And it is easy to note that ABC, AB - follow one after another. I.e. if  event A it is possible to state with high probability approach BC and so on. Events Z, F, K - do not meet at all. G> for each event there is an approach date. How such tasks dare? Circuits ?  neural networks? G> to Vprintsipe on answers for the task above - I will be already grateful. But actually the task is more difficult. To each event at its approach the number from 1 to 0 - fractional is compared. And I need to predict this number. Normally tasks such do not dare. Because it is a lot of unknown persons. Normally to be under construction model which gives adequate accuracy of a prediction. For model creation it is necessary to understand: 1) whether only events Depend on the previous events or not? If yes, whether that are identical weight? 2) whether probabilities in time Change or not? 3) whether there is inside latent process having more of states, than observable events? If probabilities in time do not change, and following event depends only from previous, model - a circuit . The matrix of approach of events is stood, in any point it is possible to predict following events and probability of a chain. If depends on several previous - that in lines of a matrix will be not one previous event, and sequence. If not equiprobablly depends from previous it is necessary to select weight. If inside more difficult mechanism already it is necessary to use the Latent Markov Model. If events concern the order it is possible to use methods of the technical analysis from trading.

3

Re: Forecasting of sequence of events

Hello, gandjustas, you wrote: G> Normally tasks such do not dare. Because it is a lot of unknown persons. G> normally to be under construction the model which gives adequate accuracy of a prediction. G> for model creation it is necessary to understand: G> 1) whether only events Depend on the previous events or not? If yes, whether that are identical weight? G> 2) whether probabilities in time Change or not? G> 3) whether there is inside latent process having more of states, than observable events? G> if probabilities in time do not change, and following event depends only from previous, model - a circuit . The matrix of approach of events is stood, in any point it is possible to predict following events and probability of a chain. G> if depends on several previous - that in lines of a matrix will be not one previous event, and sequence. G> if not equiprobablly depends from previous it is necessary to select weight. G> If inside more difficult mechanism already it is necessary to use the Latent Markov Model. G> if events concern the order it is possible to use methods of the technical analysis from trading. Similar I need to explain. The task costs in attempt to predict expenditures on a bank card on a history basis. For example the salary, the rent - with them it is easy, they are cyclic. Circuits  probably at all do not approach, well what dependence between the rent and a campaign in ? if to look as at a time series with  a type component holta-uintersa at that of these components should be a little. I such do not know. And besides budget restriction if I spent more than the certain total  I will call a taxi will not be considered.

4

Re: Forecasting of sequence of events

Hello, Gattaka, you wrote: G> Similar I need to explain. The task costs in attempt to predict expenditures on a bank card on a history basis. For example the salary, the rent - with them it is easy, they are cyclic. Circuits  probably at all do not approach, well what dependence between the rent and a campaign in ? if to look as at a time series with  a type component holta-uintersa at that of these components should be a little. I such do not know. And besides budget restriction if I spent more than the certain total  I will call a taxi will not be considered. It will be not stationary process, probabilities in time change. The external events influencing expenditures in model to put it does not turn out from a word generally. 1) It is necessary to approach on the other hand - what result is necessary to you? To predict the total of expenditures it is possible in another way, the normal analysis of time series -  the schedule of expenditures on a trend, cyclical, seasonal and casual oscillations. 2) to Remove the superfluous. Probably you all events, and specific interest not. Then removing superfluous it is possible to return to model from 1 3) After expansion abreast to add business rules which make casual oscillations by more predicted.

5

Re: Forecasting of sequence of events

Hello, gandjustas, you wrote: Under the analysis of time series there is a question. Let we predict expenditures and we have a model describing expenditures of some mass of people, i.e. it precisely predicts that people spend under new year more, it is less and so forth in the summer. As to me to use this general model for a prediction on the specific person. I.e. it is necessary to consider the general trend or some trends and singularities of the person.

6

Re: Forecasting of sequence of events

Hello, Gattaka, you wrote: G> Colleagues, G> It is necessary to predict sequence of events in time. We assume to eat events ' A ', ' B ', ' a C ' etc. the Latin alphabet. G> these events happen in time so that chain G> G> ABCDABDSSABCSSABCSS G> To me this chain is formed it is necessary to continue. As you can see in an example of all some events A, B, a C, D, SS. And it is easy to note that ABC, AB - follow one after another. I.e. if  event A it is possible to state with high probability approach BC and so on. Events Z, F, K - do not meet at all. G> for each event there is an approach date. How such tasks dare? Circuits ?  neural networks? G> to Vprintsipe on answers for the task above - I will be already grateful. But actually the task is more difficult. To each event at its approach the number from 1 to 0 - fractional is compared. And I need to predict this number. If data for "training" much, it is possible to look towards language models (language models). The most simple variant - to construct statistics n-gram to any order and to predict on its basis. Or to make language model on the basis of the recurrence neural network. There are packets: SRILM - for n-grammnyh models, RNNLM and char-rnn for recurrence.

7

Re: Forecasting of sequence of events

Hello, Gattaka, you wrote: G> Hello, gandjustas, you wrote: G> Under the analysis of time series there is a question. Let we predict expenditures and we have a model describing expenditures of some mass of people, i.e. it precisely predicts that people spend under new year more, it is less and so forth in the summer. As to me to use this general model for a prediction on the specific person. I.e. it is necessary to consider the general trend or some trends and singularities of the person. From the general model it is possible to receive seasonal coefficients, on an extreme measure the initial. And on a trend and seasonal prevalence it is necessary to do expansion on each person.