ML Models:
ML Usecase: Forecasting
<![CDATA[
Observations-
- ARIMA(SARIMA/AUTO-ARIMA/SARIMAX) Forecasting curves/results are turning flat,if trying to forecast results more than
5 days(forecast).As these models are using rolling Simple Moving Average(SMA).While Holt Winter Algo is using rolling
Exponential Average(EMA) which is tracking close with original values instead of SMA.
Though the markets are seasonal and trending but while seasonality calculation a fixed value cannot be assisgned which
will determine "α" as markets are dynamic in nature and seasonality,trends are uncertain in nature, so chances of getting
"fasle positive" Forecasting results are very high,same is seen while with actual and forecasted results.
- Another Algo are more wrapped inside standard libraries (single line of code) like KATS, AUTO-TS, LSTM,DARTS,GREY KITE.
Some give near about correct predictions but complexity/performance parameters are poor,required high GPU/CPU power compared
to normal while runnig it one by one for R&D work.
Step's Pending for Testing-
- Parallel Processing will be one option to execute 5 ML algo's on Normal Machine to get executed below 2-min without any processing
issue(High RAM).As more than 100's instruments needs to be processed in real time on hourly/daily/weekly/monthly timeframe.
- Monte Carlo(MC) simulations will be performed to get best possible forecasted value out of 10 of Forecasted curves/results, along
with 1SD and other statistical parameters though MC will generate different forecasted curves on each run for same timeframe while
other algo's forecasted results will not changed with respect to same timeframe's.
- As lot of heavy lifting needs to be performed while building ML Forecasting Algo consist of Multiple ML Models embedded in it for
all scripts for Equity,futures,Currency and Derivatives. Currently putting on HOLD as plate is already full with some other task
can be implemented in dev env.ML is still in at R&D stage.
- Thoughts/Suggestion on ML Models are always welcome.Currently looking for models which can forecast for small timeframes for
example-forecast for next 7 days.
]]>
Holt Winter Algorithm Nifty 50 Forecasting Chart-
HW & MC Forecasting Results:
<![CDATA[
- A Monte Carlo simulation is a model used to predict the probability of a variety of outcomes when the potential
for random variables are present.
- It help to explain the impact of risk and uncertainty in prediction and forecasting models.A Monte Carlo simulation
requires assigning multiple values to an uncertain variable to achieve multiple results and then averaging the results
to obtain an estimate.
- It assume perfectly efficient markets.
- In NF-50 Forecasting use-case,MC publishing 10 possible Forecasting results based on Trained data and final Forecasted
result will "mean" of it.
- Another use-case with MC is the geometric Brownian motion (GBM), which is technically a Markov process. This means the
stock price follows a random walk and is consistent with (at the very least) the weak form of the efficient market
hypothesis (EMH)—past price information is already incorporated, and the next price movement is "conditionally independent"
of past price movements.
]]>
Holt Winter Algorithm Nifty 50 Training, Testing and Forecasting Chart-
KATS Forecasting ALGO with LSTM for NIFTY 50
KATS Forecasting ALGO with LSTM for USD/INR Currency Pair
As Per KATS Model USD/INR value will be crossing 80 rupees in upcoming months due to geo-political tensions and rising inflation
and intrest rates along the global recession fear.
Though the forecasted results are quite unusual, may be KATS Algo is not exact future values but able to forecast upcoming trend.
Ref link- Visit www.livemint.com/news/india/rupee-may-slide-to-80-per-us-dollar-by-june-end
Python Notebooks:
Monte Carlo Algo on INFOSYS: infy_mc_algo.ipynb
KATS,LSTM Forecasting Algo: kats_lstm_forecasting_algo.ipynb
Arima Forecasting Algo: arima_forecasting_algo.ipynb
Monte Carlo Algo on NIFTY: nifty_mc_algo.ipynb