Approaches to Forecasting Stock Closes from Daily Price Bars

Summary

The document asks how to forecast the next day’s closing prices for S&P stocks using information from prior days. The author first considers fitting a lognormal distribution from a price bar and sampling possible closes with Monte Carlo simulation. The question is then reframed as a linear relationship between the close and the same day’s open, high, and low, with each input forecast by its own ARIMA time series and ridge regression used to handle potential multicollinearity among them.

The sole reply briefly suggests stochastic-volatility models such as ARCH or GARCH, as well as fractal and Markov-chain approaches. It does not compare these methods, explain how to estimate them, or report forecast results. There is also a timing ambiguity: using a day’s high and low to predict that day’s close would require those values to be known before the close, which may not hold for a next-day forecast. The document is therefore a collection of modeling proposals and cautions rather than evidence that any approach is suitable or predictive.

Key ideas

• The author considers using a lognormal model and Monte Carlo sampling to forecast closes.
• A revised proposal relates the close to the open, high, and low, with ARIMA forecasts and ridge regression.
• The response names ARCH or GARCH, fractal models, and Markov chains as alternative approaches.
• The document offers no comparative evidence or validation of the suggested models.
• A forecast must use only price-bar information available at the intended prediction time.

Tags

# Building predictive model for closing price using only previous days data


# Building predictive model for closing price using only previous days data












I am trying to determine which quantitative model to try and build a predictive model for the next day's closing price for all the S&P stocks based on their bar for that particular day. However, I am not sure how to think about this.

I've previously used historical data to try and predict closing prices, which would be far easier. Over here, I am thinking I can assume the stocks follow a lognormal distribution and I can then try and determine the parameters of this distribution based off of the bar and eventually use a Monte Carlo Simulation, where I randomly sample from this Lognormal to get the closing price for each stock. Does this sound okay?

Edit: I've rethought the problem and now approaching it as follows:

I am trying to find closing price using the equation below: $ClosingPrice_{t}=alpha1∗OpeningPrice_{t}+alpha2∗HighPrice_{t}+alpha3∗LowPrice_{t}$

where each of the variables opening price, the Highest value and lowest value are modelled by their own ARIMA Time-Series, as is commonly used to model stock prices. However, for the model to determine closing price I plan on using a Ride Regression as I suspect there will be multicollinearity between some of these variables.

Does this model make sense?

Thanks

## Answer by Tiberiu (score -5)

https://quant.stackexchange.com/a/21383

Yes sounds OK!

Try also stochastic volatility models like ARCH or GARCH.

Another way of forecasting would be using a Mandelbrotian fractal.

Other solution: Markov Chains

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.