TomorrowsPrices Release Notes.

© Chemcept Ltd. and Sherwood Systems Group.

TomorrowsPrices offers two major distinct tools:

1) Predictions of extent of price movement
2) Predictions of direction of price movement

Both of these basic tools have been improved and updated. In addition, two new indicative trading strategies give more realistic performance measures, and the option of HTML output is introduced. We are indebted to the hundreds of users of TomorrowsPrices for providing the feedback that has stimulated the new release. In offering a free service based on a completely new approach to stock market price movement analysis, we were aware that the initial release would have many shortcomings. We believe that this new release will be of real benefit to traders. These notes summarize the important changes made.

1) The major tool that the first release offered was to predict the extent of price movement. Despite the fact that the direction of movement was poorly predicted, knowledge of the extent is of value. The first release derived statistical information from a universal model that applied equally to all tradable assets. The model assumed that, after scaling, all assets showed the identical statistics. For the assets we studied, this assumption was adequate. However, we have now found a number of assets that show more sharply defined movements (smaller spread between 25% probability and 75% probability). Similarly, some assets have less sharply defined movements. In the new release, we have introduced a second parameter in the distribution function to capture this behaviour. This “spread coefficient” is easily estimated from as little as two weeks of O/H/L/C data. The new correlations provide a marginally improved estimate of movement given only the opening price. They provide a much-improved estimate if the first turning-point (either high or low) for the day is known. For opening price only, there are now few series we for which the median price movement is over-predicted. Under-prediction ranges up to 30% of the movement from the opening price. This precision is adequate for judging opening positions in most trading strategies. The corresponding forecast when additional, within-day prices are known is also improved. Where the first turning point is known, and we wish to predict the movement to the corresponding second turning point (either low or high), the new predictions are much improved. Very few series show an error in predicted median movement of more than 5%. As a proportion of the actual price, the corresponding error is rarely more than 0.1%. We give statistics that enable the user to judge typical performance on their series. This precision is adequate for judging closing positions for most trading strategies.

2) Direction movement is forecast in three steps. The first step is to obtain an a-priori forecast from previous day price movements. The second step is to refine this forecast on the basis of within-day movement patterns, and the final step is to refine the forecast further on the basis of the high/low range so far seen in the day. The first release forecast the a-priori direction of movement one day ahead using a pattern-matching method. This method rarely forecast directions with more than 55% confidence, and was inadequate in most cases. It uses a large set of “model” patterns to which current patterns are scaled and matched. The statistical significance of any pattern-matching algorithm is limited because there are a large number of fitted parameters. We have replaced this algorithm in its entirety with an algorithm derived from knowledge of the bimodal price-movement statistics. This method is based on pure theory with no parameters fitted from studies of actual price series. Only 4 coefficients are fitted from data. The theory is found to apply well to about 1/3 of tradable assets studied. It gives forecasts with up to 80% confidence of price movement that are well matched by actual directions of movement. In about 1/3 of series, the forecasts are adequate; the actual directions do not correspond to theory, but are still accurate at the 55% to 60% levels. This level of confidence is adequate to consistently return a trading profit. In about 1/3 of the series, the forecasts are little better than random. Tools are provided for the user to judge whether any particular series is well forecast. We do not yet understand why theory applies almost perfectly in some series and hardly at all in others. We are working on improving the price-movement model to cover a wider range of series. It should be emphasized that the extent of price movement is still well predicted even for those series in which direction prediction is poor. The within-day refinement is a new innovation in this release. Its theoretical basis is limited. However, when 2 within-day prices are available, as well as the open price, it takes account of the latest price to bias the predicted direction. The final refinement is well based on accurately established theory. Where a high to low range for the day so far is known, we know that the current day will not be a day in which either the high or the low is within the range. A rigorous refinement using the well-tested extent-of-movement statistics enables the direction prediction to be refined. As the day progresses, direction predictions become more confident, even when the a-priori predictions are poor.

3) The first release gave a profit estimate based on a possible trading strategy. However, the prediction was misleading for two reasons. First, we had confused the 25% and 75% confidence levels so that it almost always made a loss. Secondly, we did not subtract the losses to give a net profit. We have replaced this strategy in its entirety with two new strategies that are nearer to practical strategies. One strategy employs the direction prediction forecasts and the other does not. In neither case do we claim that actual net profits will match forecast net profits. The models are intended to provide an indication of whether the direction forecasts are adequate for use in a trading strategy. In the first strategy, the trader buys or sells at the opening price depending on the predicted direction. The trade is closed either when a profit level is reached, or a stop-loss level is reached. The profit and stop-loss levels are set so that, for random direction forecasts, negligible net profit is returned. In the second strategy, the trading position is not opened until the price has moved so far that we are confidently beyond the first turning-point. The position is closed when the price ceases to continue moving. This strategy loses a significant potential profit margin (the range from the open price, or first turning-point, to the price at which the trade is entered), but does not require a-priori direction forecasts. If the first strategy returns a better profit, directions are well forecast. If the second strategy returns a better profit, directions are not well forecast. Traders familiar with their series can devise better strategies then either of the alternatives discussed here.

4) Scrip issues and share consolidation. Where there was a sudden jump in price, the previous release recorded a large daily movement and gave poor forecasts, often for weeks following the event. The new release detects and corrects for scrip issues and (much less common) share consolidations giving a jump in price. Forecast accuracy is now unaffected by such scrip issues.

5) To provide a clearer output, more easily exported, we now give the option of an HTML output of forecasts.