Stock Forecast: Daily Stock Selection Based On a Self-Learning Algorithm September 25th, 2016

Stock Forecast Stock Forecast: Short-term Trading

Daily Stock Selection Based On a Self-Learning Algorithm


Time Period: January 7th, 2016 – August 31st, 2016

Daily Trading Model

The short term signals of I Know First’s proprietary algorithm can be successfully utilized besides their application for better timing of the mid and long term investments. Variety of rules based on those can be developed for trades execution and rebalancing on a daily basis.
A high predictability level and signal strength are key factors for the most intuitive approach of selecting the highest ranked stocks. For the surface plot below 1 day signals and predictabilities of the short-term model are used. The average of the daily close-to-close returns of the stock positions, as suggested by the signal direction, are depicted depending on the absolute predictability level as well as the signal strength quantiles of the respective days. The returns are higher for higher predictability and signal levels*.

Stock Forecast

*Note, that for the graph above the return averages are calculated for stocks where both the predictability level is above a given threshold as well as the stock’s signal strength is above the specified signal strength quantile on the respective day – see corresponding axes. Hence the intersection of both criteria doesn’t guarantee a certain minimum amount of trades per day.

Except the most intuitive ranking process, driven only by the raw signal strength and predictability level, there can be several ways of integrating a trend or mean reversion logic into the selection process to account for trader’s approach and/or market conditions.

The following back test results of five strategies in this context are given for the S&P 500 stocks universe since the begin of 2016. At most 20 highest ranked stocks per day (if available) are traded in each case, the equity lines represent the value of corresponding equally weighted and daily rebalanced baskets of stocks, set up to outperform the broader universe. For three of them, additionally to the predictability and signal level, the price and signal dynamics are taken into account for the respective selection processes. No technical analysis elements (indicators, oscillators) are part of the analysis below.

The signals are generated daily before the market opens and subsequently used to rank the stocks. For simplification purposes the simulation uses close-to-close price changes only and hence no limit orders or stop losses are considered for further performance enhancement. Both long and short positions can be taken, no leverage is applied. For more realistic back tests, trading only after the market opens and including the commissions and slippage, please see

The overall return in the period January 7th 2016 – August 31st 2016 ranges between 22.9% and 49.1% while the S&P 500 increased by 9.1%.


Stock Forecast

The following table summarizes the overall results and the annualized figures (assuming 252 business days in a year) for each of the strategies.

Stock Forecast

The table below breaks down the analysis into the respective trade statistics over the considered period:



Overall, without the default predictability filter applied and without considering any specific strategy, below are the averages of the daily trade returns depending on the one-day signal strength vs. the realized average S&P500 stocks return:


Focusing on the trades with a higher level of predictability further improves the returns for stronger signals:


Here, strongest 5% guarantee >=10 trades a day.

Trading Activity and Portfolio Turnover

Due to the market being mostly positive during the Jan-Aug 2016 period, most of the positions are long and constitute between 75%-90% of the holdings depending on the strategy. The average holding period per position is around 1-3 trading days, resulting in a daily portfolio turnover rate between 35% and 95%, depending on the selection process.

To understand more on the trading activity and portfolio turnover for the strategies based on the proposed daily selection processes, see the following table:

Stock Forecast

Trading activity/TV can be diminished if focusing more on selecting the most predictable assets by the predictability level. Here the average holding period can reach up to 4-5 days since predictability level is more robust than the signal, however with lower returns. Also a later mentioned “5-in-1” buy & hold tactic can be applied to hold each position for about 2.5 weeks (see next section).

Short-term Trading: Summary of Realistic Back Tests Based on Daily Stock Selection

In this section a shift from the daily close-to-close back test is done towards a more realistic market environment.

Hereby, three main features are taken into account: trading possible only after the market opens (i.e. Open-to-Close or Open-to-Open), commissions and slippage are added. The effect of those adjustments as well as the respective more realistic back-tests are depicted in the table below, for the period since launching the short-term model until Mid of July.

The commissions used represent the worst case per share commissions of Interactive Brokers according to their Tier based model – Tier 1 with 0.0035 USD per share. Only market orders are used, hence no rebates on placing non-marketable liquidity providing limit orders are distributed.

For the slippage a default model on Quantopian (open source back testing/trading platform) has been applied, which limits the possible trade volume to 2.5% of the total volume per bar (minute and day), hence splitting and delaying the order filling if the trade volume is higher. Furthermore, every minute when reaching the limit, the performance is penalized by setting the execution price 0.625 bps away from the market price of that minute.

Stock Forecast


The strategies in the table above show relatively low market exposure with beta between 0.2 and 0.6 (except the 5 days buy & hold strategy Nr. 5). Additionally, all of them generate a great positive alpha and keep the Sharpe ratios high, offering a great opportunity to generate additional high risk-adjusted returns, while further diversifying the portfolio.

The strategy nr. 5 is essentially a 5 day buy & hold strategy, whereby every day 1/5th of the portfolio is rebalanced and hold for the next 5 trading days, to avoid the “timing luck” when picking the starting date. The average holding period per position is ca. 2.5 weeks with 40% of the positions being closed after a week.

Generally, the switch to Open-to-Close or Open-to-Open trading can have a substantial impact on the overall return of some of the strategies compared to the Close-to-Close simulation. Also commissions and slippage diminish the effective returns.

However, modifying the original rules using not only the 1 day but also 2,3,4,5 day predictions allows to make strategies more robust, especially for the Open-to-Open case. For example, the strategy number 4 in the table above, which is a modification of number 1 with similar Close-to-Close results, is much more robust on Open-to-Open yielding 47.2% including commissions and slippage.

Also certain intraday adjustments such as profit taking and additional order management rules can significantly improve the performance. Below an example for the strategy Nr. 2 of the table, where the mean-reverting (bounce-back) selection logic is consistently extended to the intraday level for the period until 15th of July. Here several hours after opening all the positions, the most profitable ones are closed and the position size of the losing trades is increased in order to gain more from the expected reversion.



I Know First Research

[email protected]