- Stages now have _names and _description defined in init
- log values now passed in when init of stage; hence baseystem.init is much cleaner
- Caching:
- Cache is now accessed via a seperate object in system; so system.cache.* rather than system.* for cache methods
- Caching now done through decorators: from systems.system_cache import input, dont_cache, diagnostic, output
- Use protected=True and/or not_cached=True within decorators
- Got rid of 'switching' stages for estimating forecast scalars, forecast weights, instrument weights.
- Explicit import of a Fixed or Estimated version of a class won't work; use the generic version.
- Added seperate fields to .yaml file to switch between IDM and FDM estimation or fixed values
- Split ultra-massive accounts.py into multiple files and classes
- Split unwieldy ForecastCombine into several classes
- Added a bunch more unit tests as I went through the above refactoring exercise
- some refactoring of optimisation code - more to come
- fixed up examples and documentation accordingly
- Now requires pandas version > 0.19.0
- Capital correction now works. New methods: system.accounts.capital_multiplier, system.accounts.portfolio_with_multiplier, system.portfolio.get_actual_positon, system.portfolio.get_actual_buffers_with_position, system.accounts.get_buffered_position_with_multiplier. See this blog post and the guide
- Smooth fixed weights as well as variable: removed ewma_span and moved to new config item forecast_weight_ewma_span and same for instruments. Removed override of get_instrument_weights, get_forecast_weights method from estimated classes.
- Added extra methods to support capital scaling, but not implemented yet.
- fixed couple of bugs in getting subsystem p&l to calculate instrument weights
- removed aligned fx method, doesn't speed up and adds complexity
- solved issue #16
- Included option to show account curves as cumulative (compounding): somecurve.cumulative()
- removed percentage options, now a method for account curves: somecurve.percent()
- Incorporated capital into account curves: anycurve.capital
- General clean up of the way capital dealt with in accounting
- More speed up, couple of tweaks...
- Split up optimiser class so can selectively check if need data for equal weights; speed up
- Fixed bugs introduced in last version
- Refactored optimisation with costs code, changed configuration slightly (read this revised blog post for more )
- Introduced method to cope with pooling on both costs and gross returns, so doesn't recalculate several times
- Moved pre-screening for expensive assets to an earlier stage
- New optimisation method "equal_weights" for equal weights; means that eg expensive forecasts can be removed and then take an equal weight on the rest
- Optimisation:
- Replaced slow divide, multiply methods in syscore.pdutils with straightforward division; also means:
- Replaced Tx1 pd.DataFrames with pd.Series except where stricly necessary
- Removed a lot of defensive reindexing code where things should already be on same timestamp
- Replaced remaining reindexing code with pandas native .align methods
- accounting p&l doesn't have to work out trades, then go back to positions, if no trades provided.
- Replaced slow divide, multiply methods in syscore.pdutils with straightforward division; also means:
- Changed / added the following methods to
system.accounts:pandl_for_instrument_forecast_weighted,pandl_for_trading_rule_weighted,pandl_for_all_trading_rules,pandl_for_trading_rule,pandl_for_trading_rule_unweighted,pandl_for_all_trading_rules_unweightedSee [/docs/userguide.md#weighted_acg] for more detail. - Added
get_capital_in_rule,get_instrument_forecast_scaling_factorto help calculate these. - fixed error in user guide
- Fixed small bug with shrinkage
- Added references to blog post on costs
- introduced methods for optimisation with costs. See this blog post for more
- made a lot of tweaks to optimisation code; mainly shrinkage now shrinks towards target Sharpe ratio, equalising SR does the same; consistent annualisation
- introduced new parameter for optimisation
ann_target_SR system.combForecast.calculation_of_raw_forecast_weights(estimated version) no longer stores nested weights.
- ability to pickle and unpickle cache (
system.pickle_cache,system.unpickle_cache) - included breakout rule (example is being written)
- seperate out weighting calculation so instrument forecast pandl can be cached
- csv data is now daily and updated to present day
- Fixed bug with loading data from private module
- Changed raw cost data so returns dict not tuple
- Added 'flags' to cache identifier to replace horrors like 'portfolio__percentageTdelayfillTroundpositionsT'
- p&l for trading rules now nested in caches rather than using special identifier
- Added method
accounts.pandl_for_instrument_rules
- Renamed method
accounts.pandl_for_instrument_rulestopandl_for_instrument_rules.unweighted - Fixed bug with portfolio and instrument account curves overstating costs by adding cost weightings
- Fixed weighting of account curves and introduced explicit flag for weighting
- Added
pandl_for_trading_rule_unweightedmethod to accounts object.
- Added
pandl_for_trading_rulemethod to accounts object.
- Added t_test method to
accountCurve(and all that inherit from her)
- Added methods to accountCurveGroup.get_stats(): .mean(), .std(), .tstat(), .pvalue()
- Added method to accountCurveGroup stack; stack object can also produce bootstrap
- Added account_test(ac1, ac2) to produce a t-test statistic for any two account curve like objects.
- dynamically change class depending on config flag to estimate parameters or not
- add stage description field, and stage.methods() method
- add stage name to cache reference, always pass stage to caching function. Added cache methods to system which understand stages
- Correlation tests failing - fixed up
- Costs SR didn't get turnover - duh! Now fixed. Added a bunch of input methods to accounts object to calculate them
- tweak to account curve grouping to data frame to remove nans
- cost calculation no longer fails if no trades for an instrument
- changed buffering rounding so consistent with my own system
- Introduced maximum cap on IDM and FDM of 2.5, as per the book.
- Correlation cleaning wasn't working as documented - now does.
- cleaning up:
- renamed misleading 'get_daily_price' method to 'get_raw_price', to fix some tests that hadn't realised the difference
- fixed a bunch of tests
- changed the way cross rates are calculated to ensure data isn't lost
- Include buffering / position intertia
- Included cost data and calculations.
- Account curve improvements: generate lists of simulated trades, extend the
accountCurveobject to handle multiple columns, statistics over different periods.
- Fixed bug with bootstrapping with missing values
- Changed clean correlations so it replaces with an average
- Fixed some documentation SNAFU's
- Added reference to latest blog post
- Calculating forecast weights in ForecastCombineEstimated
- Created PortfoliosEstimated
- Calculating instrument weights
- Calculating instrument diversification multiplier
- Added a logging function
- Modified system.get_instruments so will check config.instruments (useful if estimating instrument weights)
- Included daily_prices method in data; raw data method just points to it; replaced most uses of (intraday) data.get_instrument_price with daily prices
- Added some new methods to account stage
- Cleaned up the way pooling works in correlation estimation
- Finished clean_correlation function so now deals with incomplete matricies
- Changed the way defaults feed into config objects
- Added estimation of forecast diversification multiplier to ForecastCombineEstimated
- Changed default forecast correlation estimation period; had to fix up some test output
- Changed way that forecast correlations are cached
- Started using more logical version numbering scheme :-)
- Created ForecastCombineEstimated, with get_forecast_correlation_matrices
- Added get_trading_rule_list and get_all_forecasts to forecast_combine
- Added rule_variations config option
- Added Bund data to test suite; had to fix some tests
- Pooling for forecast scalar doesn't need its own function anymore
- Changed the way config defaults are handled
- Fixed bugs: use of bool to convert str
- Fixed bugs: some test configs had wrong trading rule parameter setup; had to fix slew of tests as a result
- Added rolling estimate of forecast scalars; try System([rawdata, rules, ForecastScaleCapEstimated()], data, config)
- Moved .get_instrument_list from portfolio object to parent system
- Basic backtesting enviroment with example futures data.
- none are known
-
Simulation:
- add other trading rules (some in private...?) - cross sectional carry
- quandl data
- stitch futures contracts
- add new data from unstitched contracts (with explanatory post, include explanation for Nth contract stitching)
- Create live config from a system object (Put final value of estimates into a yaml file)
- database data
- Exogenous risk model
- check systems have correct attributes; check turnover, minimum size, right forecast scalars (distribution across instruments) etc
-
Live trading:
- ib broker interface
- accounting
- order / position reconciliation
- issue market order
- execution algos
- control functions
- get pricing data system
- Reporting:
- risk report
- risk by asset class
- interrogate signal object generated at run time
- p&l report
- trades report
- risk report