I read amongst involvement an older newspaper "Can Markov Switching Models Predict Excess Foreign Exchange Returns?" yesteryear Dueker as well as Neely of the Federal Reserve Bank of St. Louis. I direct maintain a fondness for hidden Markov models because of its dandy success inward spoken language recognition applications, exactly I confess that I direct maintain never been able to create a HMM model that outperforms elementary technical indicators. I blame that both on my ain lack of inventiveness equally good equally the fact that HMM tend to direct maintain equally good many parameters that postulate to last fitted to historical data, which makes it vulnerable to information snooping bias. Hence I approached this newspaper amongst the dandy promise that experts tin laissez passer on notice instruct me how to apply HMM properly to finance.
The objective of the model is simple: to predict the excess supply of an central charge per unit of measurement over an 8-day period. (Excess supply inward this context is measured yesteryear the % modify inward the central charge per unit of measurement minus the involvement charge per unit of measurement differential between the base of operations as well as quote currencies of the currency pair.) If the expected excess supply is higher than a threshold (called "filter" inward the paper), as well as thence acquire long. If it is lower than or thence other threshold, acquire short. Even though the prediction is on a 8-day return, the trading determination is made daily.
The excess supply is assumed to direct maintain a 3-parameter student-t distribution. The iii parameters are the mean, the marking of freedom, as well as the scale. The scale parameter (which controls the variance) tin laissez passer on notice switch betwixt a high as well as depression value based on a Markov model. The marking of liberty (which controls the kurtosis, a.k.a. "thickness of the tails") tin laissez passer on notice also switch betwixt ii values based on or thence other Markov model. The hateful is linearly theme on the values assumed yesteryear the marking of liberty as well as the scale equally good equally or thence other Markov variable that switches betwixt ii values. Hence the hateful tin laissez passer on notice assume viii distinct values. The iii Markov models are independent. The student-t distribution is to a greater extent than appropriate for the modelling fiscal returns than normal distribution because of the allowance for heavy tails. The authors also believe that this model captures the switch betwixt periods of high as well as depression volatility, amongst the consequent modify of preference (=different hateful returns) for "safe" versus "risky" currencies, a phenomenon well-demonstrated inward the menses betwixt August 2011 to Jan 2012.
The parameters of the Markov models as well as the student-t distributions are estimated inward the in-sample menses (1974-1981) for each currency dyad inward social club to minimize the cumulative departure of the excess returns from zero. There are a full of fourteen parameters to last thence estimated. After these estimations, nosotros direct maintain to also approximate the ii trading thresholds yesteryear maximizing the in-sample supply of the trading strategy, assuming a transaction costs of 10 footing betoken per trade.
With this large release (16 inward total) of parameters, I dread to come across the out-of-sample (1982-2005) results. Amazing, these are far ameliorate than I expected: the annualized returns attain from 1.1% to 7.5% for iv major currency pairs. The Sharpe ratios are non equally impressive: they attain from 0.11 to 0.71. Of course, when researchers study out-of-sample results, 1 should accept that amongst a grain of salt. If the out-of-sample results weren't good, they wouldn't last reporting them, as well as they would direct maintain kept changing the underlying model until goodness "out-of-sample" results are obtained! So it is actually upward to us to implement this model, apply it to information afterwards 2005 as well as to to a greater extent than currency pairs, to honour out if in that location is actually something here. In fact, this is the argue why I prefer to read older papers - to permit for the possibility of truthful out-of-sample tests immediately.
What exercise you lot scream back tin laissez passer on notice last done to improve this model? I suspect that equally a origin step, 1 tin laissez passer on notice come across whether the estimated Markov states stand upward for reasonably to what traders scream back of equally risk-on vs risk-off regimes. If they do, as well as thence regardless of the usage of this model equally a signal generator, it tin laissez passer on notice at to the lowest degree generate good risk indicators. If not, as well as thence perchance the hidden Markov model postulate to last replaced amongst a Markov model that is conditioned on observable indicators.
Tidak ada komentar:
Posting Komentar