CHAPTER 4-4 PREDICTION AND OTHER EXTENSION OF INFORMATION Prediction is very difficult, especially about the future. (Anonymous). [T]he most important problem which our conscious knowledge of nature should enable us to solve is the anticipation of future events. (Heinrich Hertz, "Two Systems of Mechanics", excerpted from The Principles of Mechanics, in Danto and Morgnbesser ref in pitfa45#, p. 349) The Koran says: "God forbids only what is harmful and indecent." Among the forbidden things are the drinking of alcohol and the use of intoxicants, eating of pork, gambling, usury, sexual activity outside the bond of marriage, and attempting to divine the future. ( ) Aside from the enjoyment it gives us, the purpose of getting information is to apply it to a new situation which the information can make better for us. We want better understanding of the weather so that we can arrange our actions to fit the weather that will come, and perhaps even to change the weather. We want to learn about the effects of fertilizer Z on the experimental farm in Missouri so that we can increase corn yields across the state and perhaps the world. We want to know how well the Baltimore Orioles' performance in the first half of the season will predict their performance in the second half of the season. There are some people who delight in saying that, unlike natural-scientific events, human events cannot be predicted. But it baffles me how that can be said with a straight face by any person who has ever firmly expected a friend to show up for a scheduled date, or gone to a movie whose time and place were advertised in the newspaper, baffles me. It is true that some human events cannot be predicted well -- the winner of the World Series, and tomorrow's interest rate. But some physical events cannot be predicted well, either -- which side a well-flipped coin will fall upon, or what the weather will be a year from today. Is there any important difference between the two classes of phenomena? In short, we all make predictions and act on them. And we believe that some people's predictions are better than other people's. By "prediction" I mean a statement about an event about which there will be additional relevant evidence some time after making the statement that is not available when the statement is made, information that will test the prediction. The simplest case is sports. We predict that the University of Maryland will win the basketball game tonight; after the game, we will have information that we do not now have about the winner. Weather forecasting is another obvious example. And, we may predict that there will be fewer riots next summer than last summer, or that GNP will rise by a smaller percentage this year than last year, or that copper, wheat, and oil will be decreasingly scarce in coming years. Another misconception about predicting human events -- perhaps by analogy to weather prediction -- is that the short run is easier to predict than the long run. Just the opposite is true. One can predict the direction in which the stock market will go during the next 20 years with very high surety: it will be higher than now. But one cannot predict the stock market's direction for tomorrow with accuracy much greater than chance. The same is true of the average income level, the amount of education, average life expectancy, and the price of raw materials relative to consumer goods and to the price of services. These and many other fundamental aspects of the human condition can be well-predicted for decades from now in most countries, but not for next week or next year. If the information in hand constitutes a representative sample of the entire population or collection of interest we need no additional knowledge or judgment to make a prediction. The procedure of drawing a sample randomly guarantees that, within known margins of error, the results from the sample can be generalized to the population as a whole. That is, if you randomly draw a sample of lightbulbs from your factory's production and test to see how long until each burns out, you can safely generalize to the factory's entire production with a maximum error that depends only upon the size of the sample you drew -- though of course the generalization is valid only until there is a change in the production quality, of course. All other extensions of conclusions from the observed situation to new situations, or to the future, require bringing to bear additional knowledge and judgments. Information that the Orioles baseball team added three key players since last year is relevant to an extension of the first half-season's results to the second. The angle of the sunlight probably affects the likelihood that fertilizer Z will have the same beneficial results in other places as on the experimental farm in Missouri. Prediction is a special type of generalization; it is a generalization from past to future. But unlike a generalization from a randomly-drawn sample to the population ("universe") from which it was taken, prediction is always a leap of faith; there is no scientific guarantee that the sun will come up tomorrow. Your judgment and knowledge of your subject enter into your prediction on the basis of prior experience that a particular event will occur in the future. A prediction is well-founded if it is reasonable to assume that the past and the future belong to much the same universe -- that is, if you can expect prior conditions to remain unchanged into the future. Predicting that next year's poll of architecture students about a military draft will yield much the same results as this year's poll is reasonable if conditions do not change radically. If the United States is attacked or there is peace on earth during the next year, this year's poll will probably not be a good predictor of next year's attitudes. The prediction that the sun will come up tomorrow morning is implicitly a statement that tomorrow morning will come out of much the same "universe" as have all mornings in the past. Sound predictions can be made even when changes in conditions will occur, if the changes are orderly and knowable. For example, one may predict that the amount of electric power used this year will be higher than the amount used in previous years. A shorthand way to justify such a prediction is that there has been an `upward trend' over many years. But the steady increase in power use is not a firm basis for prediction unless you know that the conditions creating the trend are also continuing. Thus, if you know that the number of generating stations, electrical appliances, and per capita income are increasing, while the price of electricity is not, you have a better basis for predicting an increase in this year's power use than if you knew just the trend. A prediction that the trend of electric power use in the 1920s would continue into the 1930s would have gone wrong because trends in income reversed due to the Depression. The past is never a guarantee of the future, and simple projection of past trends can be treacherous. An understanding of the forces that underlie the process makes it safer to extrapolate an observed trend into the future. Gaining such an understanding is a matter of saturation in the situation rather than of scientific technique. Prediction shares this property with all generalizations from the known to the unknown. There is an important difference between prediction and explanation, however. Explanation attempts to make sense of a set of data, given the available information, and hence understanding is the nub of the matter. In contrast, prediction does not rest upon understanding. Hence it is often reasonable to use mechanical rules for prediction that would not be useful in explanation; an example is the prediction that next year's crime rate will be about the same as this year's rate. From here on the discussion focuses on prediction, partly because it interests more readers and partly because the problems involved in prediction are much the same as the problems in other generalizations from the known data. The best prediction about a single entity, based on its membership in a particular category, is that the entity will be the average of that category. If all you know about a male baby is that he is American, your best prediction of his height is that he will be perhaps five feet nine inches. Similarly, if advertisement A makes 75 sales and advertisement B makes 60 sales, your best prediction for the future is that A will make 75/60 as many sales as B. If you have additional knowledge of the situation, however, you should alter your prediction to take account of it. For example, if you want to predict a male baby's adult height, there are at least two groups for reference: all males in the United States (or in the community), and his father (or parents). If his parents are six feet five inches and six feet two inches, respectively, should you predict that the boy's height will be (say) six feet four inches, or the five feet nine inches that is average for all men? In this case you probably would predict that the boy will be far taller than average, but perhaps not so tall as his parents. And you probably would make a similar prediction, though perhaps somewhat shorter, if the baby boy with the tall parents is only average length at birth. A prediction can be made on the basis of a single observation. If you have no reason to believe that the single observation is unusually high or low, then the prediction should be the same as the observation. If you do have some such outside knowledge, however, you should modify your prediction accordingly. It is quite legitimate and sensible to adjust predictions on the basis of your judgment, even if you have large numbers of observations. Much of the difference between good and not-so- good election pollsters is their art and skill in adjusting their data. The Strategy of Prediction1 This is a series of steps helps make sound predictions. Step 1: Precisely state what you want to predict. Is it the U.S. crime rate in ten years that you want to predict? If so, perhaps trends reported in the FBI Major Crime Index might be a reasonable basis for prediction. Is it the future scarcity of raw materials that you want to predict? Then perhaps the future price per unit of copper or oil or wheat, relative to wages or to the average income or to the Consumer Price Index (or not adjusted at all for inflation) is the measure of scarcity that fits your needs. Whichever, make your aim precise. Step 2: Determine the period for which you wish to forecast. Crime, or raw material scarcity, for next week? Next year? Ten years from now? Thirty years from now? Or during a specific time, such as when the Republican National Convention will be held in your city? The appropriate method for a particular prediction depends on the period in which you are interested. For example, if you want to predict how many people will be killed in auto crashes next year, simple extrapolation of the number killed this year would probably be best. If you want to predict how many will be killed in the year 2003, however, you would probably take into account design improvements that might make cars and highways safer, changes in the size of the population, and other fundamental factors. You might even consider the possibility that some entirely new mode of transportation will be invented in the meantime. (Imagine someone living in 1903 trying to predict the number of horse-related accidents just as automobiles were becoming popular.) Step 3: If you are predicting a future state of affairs, begin with the present state as your benchmark. There must be some continuity, some non-independence or similarity between contiguous events, if there is to be any better-than-chance predictability in this world. And if there is continuity, the present is the most reasonable benchmark for a prediction, all else being equal. Step 4: Determine how much variability there has been during a past period equal in length to the amount of time between now and your prediction date. If there has been no variability at all, then your task has been simplified. For example, if you want to know the likelihood that the state of Maryland will fulfill its pension promises even though it has not put the required amount aside, finding that all other states in similar situations have paid the pensions should reassure you. If you want to know the amount of liquor that people will drink ten years from now, and if liquor consumption per capita has changed very little from year to year in the past twenty years, then you can safely predict that consumption per person ten years will be what it is now. If you have reason to think that a new factor likely to alter past patterns has entered the situation -- a prohibition law, for example -- then of course you will not predict that the future will resemble the past. Otherwise, however, constancy in the past suggests that the present status is your best prediction for the future. It can be disastrous to assume wrongly there will be little variability. Based on the fertility rates of the 1930's, demographers predicted in 1947 that U.S. population in 1970 would be 140 million and going lower; instead, the population turned out be 205 million, much to the demographers' embarrassment. Step 5: Look for pronounced cyclical effects. Can you find a strong and obvious cyclical regularity in your data? Anyone who, in Toronto in August, predicts that the temperature outside will be the same six months hence as it is that day, is doomed to shiver. Plotting a few years' temperatures month-by-month on graph paper will correct this prediction. Be very reluctant, however, to conclude that there is a cycle in the phenomenon you are interested in. Purely random fluctuations often look very much like cycles. I'd hazard that nine times out of ten when people think they see a cycle -- and often are willing to bet on it -- the cycle is pure illusion. Step 6: Examine the long-term trends. Check whether there has been a systematic long-run drift. The easiest method is to compare the averages of the more distant and more recent past periods -- say, comparing average liquor consumption in the decades from 1950 to 1990. Plotting the data on a graph is a good way to examine for trends. If you find that there has indeed been a trend during the relevant period, then it may be reasonable to extend that trend for your prediction. But here you must exercise extreme caution. The biggest danger is making a long-run projection from a short- term trend, such as an upswing in liquor consumption within the last three years. A gradual decrease in liquor consumption over the past three decades is likely to be a safer basis for predicting a slight decrease next year compared to this one. One of the many sad examples of fallacious trend extension was the conclusion by individuals and governments that the raw material scarcity that appeared in 1973-1974 heralded a continued decline. The most important element in making sound predictions from trends is to grasp the long sweep of history. Obtain evidence from as far back in the past as you can, to ensure that what you think is a trend is really not just a blip in history. Why is it reasonable to assume that a trend will continue? The short answer is that continuity is a basic feature - perhaps the most basic feature - of the world we live in. If the pattern of the past is not consistent, you must decide which parts of the past to pay most attention to. It is generally reasonable to weight the recent past more heavily than the distant past, though never completely forgetting the distant past. Step 7: Find other variables that go together (though not necessarily causally) with the variables you want to predict. If you have little or no data on a phenomenon you want to predict, you may analyze change in another phenomenon that is closely related to what you are interested in, and for which data are available. The relationship between the phenomena need not be causal. For example, there are no reliable direct observations of the number of Mexican aliens illegally in the United States in various years. But there are data on the numbers of deaths of illegal Mexican aliens. Trends in deaths can be used to predict trends in living persons, even though deaths do not cause live persons. And if you have known from demographic calculations what proportions of people die each year, you can predict the trend in number of living persons even better. Step 8: Find variables that cause changes in your phenomenon, and predict changes in those variables and then in the associated changes in your phenomenon. If you know how many bicycles will be sold - or better, how many will be in use in a given year - you have an aid to predicting sales of bicycle tires, because very few bicycle tires are sold except to be used on bicycles. And perhaps the number of teenagers, together with the income level, will predict the number of bicycles sold, because teenagers buy many of the bikes sold, and income affects all purchases. Here we bringcausal influences into our prediction process. Thus, if bicycle sales are a major predictor of bicycle tire sales, then income and age will be causal variables, and those variables themselves must be predicted. Step 9: For very, very long-range predictions, consider first whether there are very long-run trends in the phenomena that will continue. The very long-run trends in liquor consumption and in murder in the United States are downward. Unless you have some strong arguments against the continuation of those trends, they may be your best basis for predictions. But if you do have one or more strong reasons to believe that the bases for the previous long- run trends, have changed, then you must attempt a wide-ranging forecast of what society will then be like, in light of the major technological and social changes that you expect will take place by then. The greater predictability of small changes explains the businessperson's willingness to try small changes in price and other variables more readily than large (and sudden) changes. This concept also fits the United States' pragmatic and meliorist philosophy of government. The underlying logic is that the effect of major changes -- that is, revolution -- is impossible to predict well because any major alteration -- toward socialism, say -- must affect not only economic variables but also the forms of justice, international policy, and so on. (More about this in Chapter 00 on "muddling through".) PREDICTION, SCIENCE, AND JUDGMENT Here is a homely example of prediction. You observe that the owner of the third house on the block owns a big new car. If you predict that the next-door neighbor in the fourth house also has a big new car, you will be right more often than if you guess that the neighbor has a small old car. Furthermore, if you then observe that the fourth house owner does have a big new car, a prediction that the fifth house owner also has a big new car is more likely to be correct than if only the fourth house and not the third house had a big new car. That is, the high degree of similarity between the third and fourth house increases the chance of accuracy in a prediction that the fifth house would be like the fourth house. The dimension on which one element is "next to" another need not be geographic. Frequently the relevant proximity is in time. The number of minutes of daylight tomorrow may be predicted on the basis of the number of minutes of daylight today and yesterday. A patient's mental state tomorrow may be predicted on the basis of her mental state today. Also, if two elements are similar on one dimension, they may be similar on another dimension as well. For example, if two students get the same score on an I.Q. test, one would predict that they would get the same grades in school. And if the same quantity of wheat is harvested in two given years, one would predict the same wheat price for the two years. (Remember, however, that they may not be similar to each other in ways that are important for your particular prediction or purpose, if the dimensions on which they are classified as the same are not relevant, as we shall see later.) One may expand the concept of sameness a bit, still with simple examples. Consider trends, now. In some cases we do not assume that C will be the same as B but that, say, C will be proportionally as much larger than B as B is large than A. So we now have two notions of sameness, the sameness of the elements and the sameness of the proportional differences between pairs of the elements. An example of predicting on the basis of such differences is founded in a crude or "naive" forecast of gross national product. If GNP has been growing at the rate of 5 percent per year in recent years, one may forecast that next year`s GNP will be 5 percent higher than the GNP recorded this year. It may also be useful to observe that elements are similar in more than one dimension, suggesting a prediction of one from the other using the information from both dimensions. If you are told that two brothers are both five feet seven inches tall, and one weighs 150 pounds, it might be reasonable to predict that the other weighs 150 pounds, also. But if you are also told that both wear size 37 suits, you would have even more basis for a prediction that the second weighs 150 pounds. And if you are told that the first is five feet seven inches, size 37, and 150 pounds while the second is five feet seven inches and size 36, you could predict the weight of the second with more confidence than if you had only his height or suit size. Projections of trends are most appropriate when the causes of past levels of the dependent variable are a) many in number, b) each so small that it cannot be identified, and c) none of the biggest of them is subject to major discontinuous change. Such a case might be the increase in the divorce rate next year, or the stock market level tomorrow; there is ample evidence that the best predictor of the level of the stock market as a whole tomorrow is the level today. But where the causes are few and relatively large - for example, the rate of growth of your personal record collection, where growth depends entirely upon your own decisions, automatic trend projection is a) inferior to causal analysis, and b) unwise if the major cause is subject to someone's control and/or may experience a discontinuous change. To clarify the relationship between the "naive" method, the causal method, and the social forecast method, let us imagine ourselves standing in 1880 trying to predict the number of horseshoes that will be sold in the future. If we want to predict the number of horseshoes that will be sold in the next week, the most reliable predictor - assuming that there is nothing special about the week - will be the number of horseshoes that were sold last week. If, on the other hand, we wish to predict the number of horseshoes that will be sold five years hence, it might be reasonable to make a causal analysis of the relationships of income and the age distribution of the population to the number of horses, and thence from the number of horses to the number of horseshoes used per year. If, however, the horizon for forecasting were 50 or 100 years (as of 1880), neither of these methods would be appropriate. Rather, it would make more sense to look around at the structure of the economy and of technology, notice the steam engine and its rise and the steamship and their rises, and guess that perhaps something will come along that will replace horsepower entirely. Lest it seems easier to backcast with twenty-twenty hindsight than to forecast, let us look forward and imagine that a prediction about sales of automobile tires a hundred years from now ought to take into account totally new materials that automobiles might ride on -- assuming there are automobiles -- and perhaps that the wheel itself may be replaced by other methods of locomotion such as air cushions under land vehicles. Making such long-term predictions is extremely risky. But a well- exercised imagination plus sensibly-gathered information about experience in the past, in combination with well-chosen theoretical ideas, often make it possible to forecast far better than chance. For example, forecasts made regularly in the past - - say, every ten or a hundred years -- that the price of iron would be lower rather than higher in the future, would have been correct much more often than they would have been wrong. Indeed I have offered to bet that this will be true for the future as well. (The offer was accepted in September, 1980, by biologist Paul Ehrlich and two of his associates, who picked four metals and the settlement date of ten years ahead in 1990.) THE BASIS FOR PREDICTION ITSELF2 Why should any prediction be any better than random chance? If you are interested in fundamental questions, read on. If you like to skip the philosophy, you can safely do so. Sitting in the social sciences teachers' lounge at Hebrew University in Jerusalem in 1968, I heard economist-statistician Yoel Haitovsky say, "God wrote a set of equations, and it is our task to find out what the equations are." He was expressing a view of science which became dominant in the 19th century. As Kant put it, the "Author of the world" caused there to be "the glorious order, beauty, and providential care everywhere displayed in nature." (1787/1965, p.31). This surely is still the common view today. I replied to Haitovsky roughly as follows, "God wrote no equations. Rather, God (though let's not get hung up on that word) gave the formless mass a kick and deformed it so that it is no longer totally random and without order. We try out various equations to see which ones come closest to fitting our needs". The view that it is scientists who invent neat models to imitate some aspects of messy reality has come into vogue because of Einstein and Bohr. Einstein was fond of saying that physical theories are the products of human imagination. On this matter, Kant pointed in the right direction, Einstein notes (1954, p.22). Kant wrote that "reason" -- by which he meant something like what we call theory or logic or both -- must "show the way with principles of judgment based upon fixed laws, constraining nature to give answer to questions of reason's own determining" (p. 20). He says that the scientist should not approach nature "in the character of a pupil who listens to everything that the teacher chooses to say, but of an appointed judge who compels the witnesses to answer questions which he has himself formulated" (p. 20). And he goes on to say that we can know of phenomena "only what we ourselves put into them" (p.23). Admittedly it is unclear whether Kant was talking of experience-based theories as we know them, or purely tautological mathematical propositions. (He claims to be talking of the latter, but he understood the former quite well, and alluded to such theory though excluding it from his "a priori" category.) In any case, Kant surely did emphasize that our scientific vision of a phenomenon is, and should be, a product of our ideas about how best to think about the phenomenon, as well as the data that we collect about the phenomenon, just as Einstein and Bohr later insisted was the heart of their scientific approaches. This subject was on my mind when I overheard Haitovsky's remark because I had recently been working on the relationship of income to family size, and I had once more seen how it made no sense to talk about a "true" or "underlying" relationship, and how it is impossible to ever "purify" a relationship of extraneous forces in such fashion as to reach a "true" model. Rather, one must always make a choice of which model to work with, and the choice must be made with reference to the purposes of the investigation; different purposes should lead to different views of a scientific relationship, unattractive as this seems because of its apparent lack of "objectivity". It also lacks attractive power, I think, because its appeal is not to a particular theory being out-and-out correct but rather to its being approximately correct, or being "convenient", in Bertrand Russell"s term (1945, p.832). This is not palatable to people who view themselves as searching for eternal truth rather than searching for workable ideas that will help us get on with additional learning about the world and with dealing with our human problems more successfully. But perhaps the imprimatur of Einstein and Bohr will have some persuasive power; they are tough names to be contemptuous of, even if they are not names that I can actually conjure with. We can illustrate this principle with the shape of the earth. Would you say that the earth is round? I have exasperated my children by saying that I mostly regard the earth as flat, as is also appropriate for carpet-layers. Geophysicists consider it an elongated sphere. Airline navigators plotting a route from Europe to the U. S. properly regard the earth as round, alright, but pilots flying from Denver to San Francisco better view it as jagged, or else they won't view it very long. Taken altogether, the earth "is" not anything, and to say that it "is" anything falls into the snare of the word "is" mentioned earlier. The earth equals only the earth, and any descriptive adjective at best grasps one aspect of the earth. In similar fashion, Benoit Mandelbrot, creator of the mathematics of fractals, discussed the coastline of Great Britain: its length depends upon whether you measure it as the crow flies, or as a beachcomber would walk it, or in some other fashion. This view of the world as deformed ex-chaos also provides a comfortable basis for prediction. Here is an example of its fruitfulness. For some time now, following Harold Barnett (Barnett and Morse, 1963)I have written about how our supplies of energy, land and other natural resources have been increasing rather than decreasing, contrary to all common sense and Malthusian diminishing returns. (This is in connection with my central interest, the economics of population.) The central evidence for this general proposition is that the prices of natural resources relative to our most precious commodity--the hours of our lives--and also even relative to consumer goods, have been decreasing rather than increasing. This has led to the prediction that this trend, which has been in operation over the entire span of history for which we can muster any data -- and that goes back almost 4000 years for copper -- will continue into the future indefinitely. Unsurprisingly, not exactly everyone agrees with this prediction. And naturally enough, some persons have inquired about the method which is the basis of this forecast. Simplifying dangerously, a trend may be expected to continue unless there is a persuasive theoretical reason to think that it will not. And in the case of resource availability such a persuasive counter- theory does not exist; in fact, a plausible theory says quite the opposite, that the problem of impending shortage mobilizes people to develop solutions that leave us better off than if the problem had not arise. Furthermore, if statistical evidence is overwhelming, one may choose to believe that a trend will continue even despite persuasive theory to the contrary, implying that the theory needs rethinking. The underlying philosophical basis for this general proposition about trend continuation is the same as the basis for any prediction, that there is likely to be a positive correlation between contiguous observations, in consonance with the idea that the world as we know departs from complete chaos, from complete randomness. Now of course there is a certain amount of pulling oneself up by one's bootstraps in this way of thinking. It says that we may predict the future on the basis of past experience because the world is non-random, and we believe that the world is non- random because we have been able to predict successfully in the past. But if we look for it, we can find that Hume supported his assumption that prediction is possible with a similar observation: "[I]f all the scenes of nature were continually shifted in such a manner, that no two events bore any resemblance to each other, but every object was entirely new, without any similitude to whatever had been seen before, we should never, in that case, have attained the least idea of necessity, or of a connexion among these objects."(1949, p. 104). (Notice that though Hume says that we can never learn the nature of the physical connection among events, he does not deny that there is a connection). And he makes a similar observation specifically with respect to social science. "[W]ere there no uniformity in human actions, and were every experiment, which we could form of this kind, irregular and anomalous, it were impossible to collect any general observations concerning mankind; and no experience, however accurately digested by reflection, would ever serve to any purpose" (p.106). ( But the idea of non-chaos due to a single deformation is a structural vision that is not identical with the success of prediction; it is itself a hypothesis about the physical nature of the world and about a historical event. To the extent that this idea is sound, it buttresses the naked observation that prediction based on experience does better than random chance. And therefore the intellectual operation is not just a bootstrap operation, not just tautology and restatement. In fact, this view fits with recent speculations by astronomers. Such physical corroboration, however, is not crucial for present purposes though it may make one feel more comfortable with the hypothesis.) SUMMARY Though prediction always involves uncertainty, there are ways to improve the accuracy of predictions, including being precise about what you want to predict, taking the past into account in a sensible - not a mechanical - manner, and finding variables consistently related to the variable you want to predict. Good prediction, especially for the medium and long-term, depends on imagination and judgment formed through long experience with your subject. AFTERNOTE Mark Twain provided this amusing commentary on forecasting: In the space of one hundred and seventy-six years the Lower Mississippi has shortened itself two hundred and forty-two miles. That is an average of a trifle over on mile and a third per year. Therefore, any calm person, who is not blind or idiotic, can see that in the Old Oolitic Silurian Period, just a million years ago next November, the Lower Mississippi River was upward of one million three hundred thousand miles long, and stuck out over the Gulf of Mexico like a fishing rod. And by the same token any [continue] frcst44# 176 July 3, 1990 FOOTNOTES 1Armstrong's (1978) practical and delightful book provides a host of practical suggestions for forecasting. 2 David Hume makes an argument for the basis of prediction not unlike the view which I offer above. Hume says, As to what may be said, that the operations of nature are independent of our thought and reasoning, I allow it, and accordingly have observed that objects bear to each other the relations of contiguity and succession; that like objects may be observed in several instances to have like relations; and that all this is independent of and antecendent to the operations of the understanding. But if we go any further and ascribe a power or necessary connection to these objects, this is what we can never observe in them, but must draw the idea of it from what we feel internally in contemplating them. And this I carry so far that I am ready to convert my present reasoning into an instance of it by a subtlety which it will not be difficult to comprehend. When any object is presented to us, it immediately conveys to the mind a lively idea of that object which is usually found to attend it, and this determination of the mind forms the necessary connection of these objects. But when we change the point of view from the objects to the perceptions, in that case the impression is to be considered as the cause and the lively idea as the effect, and their necessary connection is that new determination which we feel to pass from the idea of the one to that of the other. The uniting principle among our internal perceptions is as unintellible as that among external objects, and is not known to us any other way than by experience. Now the nature and effects of experience have been already sufficiently examined and explained. It never gives us any insight into the internal structure of operating principle of objects, but only accustoms the mind to pass from one to another.(A Treatise of Human Nature, p. 115, edition unknown.) PUT IN RES METH FOLDER For prediction and classification sections, see changes made in draft for Thinking book. now in file assurance. take out when used. People believe that they know their own situations so much better than another person can know it that they quickly reject advice and ideas about their situations. Sometimes this is well-warranted because the advice-giver really does know too little about the situation to be able to offer good advice; this is the main reason that socialist central planning fails. But sometimes the advice-giver does not enough about the situation to be able to bring to bear some more general knowledge that results in useful advice, but is rejected without sufficient hearing. "You don't understand my industry" is too often an automatic response, and often a costly one. That was the universal reaction when I proposed a volunteer auction plan to the airline industry as a way of dealing with the overbooking problem. And it is the reaction of my close friend T when I suggest to him that he do a bit of market research about what consumers like in the garments he sells; he is sure that his accumulated knowledge is sufficiently great that it is not worth spending even one thousand dollars and a bit of time in systematic investigation, though this seems never to have been done in his industry. Perhaps this assurance is crucial, and the absence of it damaging. I think that I would have been a better father if, during their adolescent years, I was more sure that I knew what I was doing, and that my decisions were good ones. Instead, my lack of confidence led me to indecision and weakness in the face of their determined arguing that we should have the same rules as other parents, and that the kids knew best what was good for them, indecision and weakness that I now believe were both unhelpful and unnecessarily painful. Even I really did not know well what I was doing, it probably would have been better if I thought that I did, and acted upon that belief. Page # thinking frcst44# 3-3-4d