Methodological issues in forecasting - Taylor & Francis Online

Journal of Economic Methodology 12:4, 517–542 December 2005

Methodological issues in forecasting: Insights from the egregious business forecast errors of late 1930 Robert S. Goldfarb, H. O. Stekler and Joel David

Abstract This paper examines some economic forecasts made in late 1930 that were intended to predict economic activity in the United States in order to shed light on several methodological issues. We document that these forecasts were extremely optimistic, predicting that the recession in the US would soon end, and that 1931 would show a recovery. These forecasts displayed egregious errors, because 1931 witnessed the largest negative growth rate for the US economy in any year in the twentieth century. A specific question is what led forecasters to make such serious and substantial empirical errors. A second more general issue involves the methodology of forecasting. The 1930 forecasts were sometimes based on explicit analogies with previous serious business cycles. Modern forecasting approaches are based on techniques that may not be recognized as analogies. Using the 1930 forecasts, we examine the implicit-analogy content of forecasts, and what might render such implicit analogies valid or invalid. This 1930 forecast example also resonates beyond the confines of economic methodology because forecasts about the Great Depression are of continuing interest to the profession at large, and we produce a forecast series not previously available. Keywords: Analogies, qualitative forecasts, great depression, forecast evaluation, business cycles

SOME BACKGROUND: THE LITERATURE ANALYZING GREAT DEPRESSION FORECASTS A number of economists have used economic forecasts to try to explain some features of the Great Depression. In some cases the underlying forecasts were actually prepared during the Depression years. In other cases forecasts were constructed after the fact or were those implied by financial variables such as interest rates or futures prices. Forecasts created after the fact sought to determine whether the Depression could have been forecast or whether that period’s deflation, especially during 1930, was anticipated. For example, Dominguez et al. (1988) used modern time series techniques to make forecasts for the early years of the Depression. They found that their VAR model could not predict the extent of the decline or the Journal of Economic Methodology ISSN 1350-178X print/ISSN 1469-9427 online #

2005 Taylor & Francis http://www.tandf.co.uk/journals DOI: 10.1080/13501780500343524

518

Articles

magnitude of the actual deflation. They concluded that the Depression was not forecastable. Other studies focused on whether the 1930 deflation had been anticipated. The results are conflicting. Hamilton (1992) examined futures prices. He found that while actual commodity prices were in fact declining, the futures market seemed to anticipate that they would increase. He concluded that the deflation in the first eight months of 1930 was not anticipated.1 Evans and Wachtel (1993) concluded from interest rate data that the deflation of the early Depression years was not anticipated. On the other hand, Cecchetti (1992) used an expected deflation measure derived from interest rates and found that the deflation could have been anticipated. The aforementioned studies used forecasts created after the fact. Only a small number of papers have investigated the ex ante real-time predictions actually made in 1930. Nelson (1991) examined business periodicals to see whether the business community recognized that deflation was taking place. He concluded that until mid-1930 only a mild deflation was anticipated, but after that a more severe deflationary period was expected.2 Similarly, Klug et al. (2002) analyzed a set of actual forecasts made by railroad shippers throughout the 1930s and concluded that these shippers consistently overestimated the demand for railroad cars. Those shippers did not believe that demand would decline as substantially as it actually did. While Klug et al. showed that simple rules of thumb, which could have been utilized in that period, would have produced better forecasts, they did not analyze the causes for the less accurate actual forecasts. In a major study using real-time forecasts made in 1929 and 1930, Romer (1990) examined the predictions of some banks and advisory services. She concluded that, after the Crash and during the early part of 1930, these groups felt that there was a great deal of uncertainty about the economy’s future direction. SOURCES OF FORECASTS: OCTOBER 1929 TO DECEMBER 1930 Examining forecasts for the US issued monthly from October 1929 to December 1930 allows us to gain perspective about forecasts prepared at the end of 1930. These forecasts come from a variety of sources. Most were obtained from the Commercial and Financial Chronicle, a weekly publication that compiled a tremendous amount of information for the business and financial community. Summaries of the forecasts of many banks were included in this publication. In addition, The New York Times regularly published summaries of the forecasts of two major New York banks and frequently published the forecasts of other banks as well as the speeches of well-known business forecasters. In some instances the same forecasts were summarized in both publications, but there were differences in emphasis. We also examined the forecasts of

1930 forecasting errors

519

the Harvard Economic Society.3 We found assessments and forecasts from thirty-two individuals and institutions; their names are listed in Appendix Table A1. EVALUATION METHODOLOGY Scoring procedure Virtually all of the forecasts of the 1930s were non-quantitative, so we cannot undertake a quantitative evaluation such as those used to analyze forecasts made in recent years. However, we have developed a procedure for evaluating them. The individual forecasts gave qualitative assessments both about the current situation and about the direction the economy would take in the near future. We paraphrased the forecast summaries from each source, and developed a scoring system which we applied to these paraphrased qualitative statements. We then constructed a time series of ‘representative’ forecasts by averaging the scores of the individual forecasters for each time period. Using a scale running from 21 to +1, with gradations of 1/4, we assigned separate scores to the assessments of current conditions and to the forecasts of future activity. For example, a statement that production and prices have stabilized would receive a score of 0; if the individual then predicted that the economy would recover gradually over the next several months, that forecast would be scored as +1/2. Other values were assigned if the forecasters’ statements contained different assessments about either current or future conditions. Thus every forecast received two scores: one for the current situation, the other for the forecast about the future. Table A2 in the Appendix describes more fully how the scores are related to qualitative statements. The forecast statements were scored separately by two of the authors. When the scores differed, we discussed our reasons for assigning the specific scores and reconciled them. We also reconciled the differences in emphasis in the forecast summaries of the Commercial and Financial Chronicle and The New York Times whenever the two sources described the same forecast. The analysis that we present is based on these ‘reconciled’ scores. Validating the scoring procedure We next validated this methodology. During the 1930s, the Federal Reserve published a monthly analysis of economic conditions in the Federal Reserve Bulletin.4 This analysis was also qualitative, and we applied the scoring procedure described above to these statements.5 We cumulated the monthly scores assigned to the FED’s assessments of the current economic situation. This creates a time series that, if our procedure is correct, measures

520

Articles

economic activity over a period of time. This constructed time series was then compared to the movements of the Federal Reserve Board Index of Industrial Production over the same period of time. Figure 1 shows that the two time series had similar movements between October 1929 and December 1930. Both showed a decline in late 1929, followed by sideways movements in early 1930 and a sharp decline later in 1930. Since our scoring system, applied to qualitative assessments of the state of the economy, replicated the economy’s movements, we are confident that our scoring system enables us to also evaluate the qualitative statements of the forecasters. Evaluation procedures The evaluation of the representative forecaster’s assessment of current conditions differs from the evaluation of that forecaster’s predictions of future conditions. It is assumed that the assessments of current conditions contained in any prediction refer to the situation that prevailed in the previous month, e.g. an analysis issued in early April refers to conditions during March. We further assume that the score we assigned to the FED’s analysis of current conditions accurately measures what actually happened in that month. Since we scored the FED’s analysis of current conditions the same way as the forecasters’ assessments of those situations, we were able to directly determine how good the forecasters were in assessing current conditions. For each month, the forecast error is the difference between the FED’s score and that of the representative forecaster. While the absolute magnitude of the error has no meaning, the sign of the average error determines the direction of the forecast bias. The method for evaluating the forecasts is slightly different because the forecasters’ scores must be compared with values assigned to the actual outcomes some months in the future. While many forecasts indicated that they referred to economic conditions expected to prevail in the next quarter, other statements were not as precise. Therefore, we explicitly assumed that the forecasts were about the state of the economy during the three months following their issuance. We compared the scores that were assigned to the forecasts in time t with the average of the scores that we assigned to the FED’s analysis of current conditions in months t, t+1, and t+2. Then, by averaging over the actual scores for that 3-month period, we are able to compare the actual and predicted values. RESULTS: OCTOBER 1929 TO DECEMBER 1930 The scores assigned to the representative forecaster’s assessment of current conditions and to the FED’s interpretation of those conditions are both


521

Figure 1 Top panel: Federal Reserve Board Index of industrial production October 1929 to December 1930. Bottom panel: Authors’ cumulative scoring index, October 1929 to December 1930 (Oct. 295100)

522

Articles

presented in Table 1. The results indicate that on average the mean assessment is biased. It is more optimistic than the FED’s analysis of the same conditions. This is especially true for the period from May 1930 onward when the US economy declined sharply. Even though the forecasters realized that the economy was still declining, they tended to underestimate the extent of this weakness. Table 2 presents an evaluation of the actual forecasts.6 With the exception of October 1929, the scores of the mean forecasts that were made prior to March 1930 do not differ substantially from the scores of the FED’s assessment of conditions that prevailed in the 3 months after the forecasts were made. These forecasts would be considered accurate. All of the forecasts made after March 1930 were grossly inaccurate. They all failed to predict that the economy would continue declining.7 It is necessary to explain the dichotomy between the analysis of current conditions that, while biased, correctly showed that the economy was still declining and the forecasts that continuously and erroneously indicated that the economy would turn up in the following 3 months. We particularly want to focus on the forecasts that were made in the last quarter of 1930. While Romer (1990) had concluded that the forecasts made in late 1929 and the early part of 1930 displayed a great deal of uncertainty about the economy’s future direction, the forecasts made at the end of 1930 do not display this uncertainty. The consensus of these predictions was that the recession would soon end. The optimistic flavor of these forecasts is illustrated by the following representative quotations: We are now near the end of the declining phase of the depression. (Harvard Economic Forecasts 15 November 1930) experience during past depressions indicates that the decline is nearing its end and that recovery will begin in the first half of next year, and probably in the first quarter. (Harvard Economic Forecasts 29 November1930) … we conclude that a business upturn … probably will be in progress by February or March 1931… (Persons 1931: 19)8 Moreover, based on our scoring system, in the fourth quarter of 1930, there were only a few forecasters who predicted that the economy would decline further. These forecasts were made in October and November. None of the December forecasts suggested that a further decline was possible. Since 1931 witnessed the largest negative growth rate for any year in the twentieth century, we must examine why these forecasts were monumentally erroneous.9 EXPLANATIONS FOR FORECAST ERRORS These forecast errors could have been caused by any of a number of factors: (1) failure to apply the prevailing economic theory; (2) misinterpretation of the

Table 1 Assessment of conditions prevailing in September 1929–November 1930 9/29 10/29 11/29 12/29 1/30 Score of Representative Forecaster Score of FED Survey Error

2.40 2.21

2/30

3/30

4/30

5/30

6/30

2.55 2.33 2.14 2.08

2.09

+.04

2.04

2.23 2.61 2.15 2.15 2.22 2.14

2.50 2.41

+.25 +.21

2.25 21.0 21.0 2.25 2.25 2.50 2.50 2.21 2.77 2.39 2.10 2.10 2.28 2.36

2.25 2.50 21.0 2.50 +.15 2.29 2.45 2.17

+.50 +.64

0 +.08

7/30

8/30

9/30 10/30 11/30

Mean Error52.163 (the average of the row 3 error from 9/29 through 11/30) Note: The assessment in row 1 of conditions prevailing in a particular month is based on the individual’s statement in the subsequent month (i.e.– a statement made in an early July report about current conditions refers to the situation prevailing in June).

Score of Representative Forecaster Avg. score of FED Survey For t, t+1, t+2

10/29 11/29 12/29 1/30

2/30

3/30

4/30

5/30

6/30

7/30

8/30

9/30

2.25 2.41 2.08 2.15

0

+.18

+.16

+.14

+.11

+.19

+.28

+.12

2.67 2.33

0

0

10/30 11/30 12/30 +.14

+.14

+.30

2.08 2.16 2.33 2.75 2.75 2.50 2.33 2.42 2.58 2.50 2.25


Table 2 Evaluation of forecasts issued in October 1929–December 1930

523

524

Articles

real-time data; (3) misapplication of the existing forecasting procedures; and (4) using forecasting procedures that, after the fact, turned out to be inappropriate. Relationship of forecasts to prevailing economic theory Were the forecasts that the Depression was coming to an end consistent with the prevailing economic theory? A thorough treatment of this question is beyond the scope of this paper, requiring as it does a full investigation of the range of business cycle theories available at the time. Moreover, the forecasts we summarize above are often those of business forecasters, so the expectation that they would be consistent with ‘received economic theory’ is somewhat unreasonable.10 Nonetheless, business forecasting books indicated that in the vicinity of the lower turning point, the data should be consistent with one belief common to the prevailing economic theories. While these theories had different explanations both for the causes of cycles and for the subsequent revival, they often were in agreement with the following idea. The trough occurs when production is less than consumption. At that point excess inventories begin to decline, and any increase in demand leads to increased production.11 A number of the qualitative forecasts that we reviewed had mentioned this relationship and asserted that the economy must be near the trough of the cycle because the decline in production had exceeded the decline in consumption. Thus, the forecast analyses were not in conflict with this feature of economic theory, but the question is whether these statements were in accord with the data. Interpreting the real-time data If production had, in fact, declined more than consumption, then inventories should have begun to decline or at least not increased. Some data were available in real-time in 1930. We use those data to examine the relationship between these variables. Although the Survey of Current Business was not as comprehensive as it is today, in the period in question it did publish a substantial amount of data on finished manufacturing goods inventories from ten major sectors: (1) foodstuffs; (2) textiles; (3) iron and steel; (4) nonferrous metals; (5) lumber; (6) stone, clay, etc; (7) leather; (8) rubber; (9) paper; and (10) chemicals and oils. Figure 2 compares the movements in those manufacturing goods inventories with movements of the Federal Reserve Board Index of Industrial Production. That data clearly show that inventories did not decline in 1930 even though production had been dramatically reduced.12 Inventories were stable or even rising. Thus, an improvement in economic conditions was certainly not evident in the inventories and production data that were available at that time. Applying this aspect of prevailing economic


525

Figure 2 Movements in industrial production and inventories US 1928 to 1930

theory to the data provides no evidence to support an end to the downturn. If the theory in question had been used and if the data had been interpreted correctly, forecasters would not have made that forecast. CONTEMPORARY DISCUSSIONS OF HOW FORECASTS WERE AND SHOULD BE MADE We now turn to discussions by contemporary forecasters of how forecasts are and should be made. A first question to consider is what heuristic rules, if any, forecasters had for identifying troughs. Heuristic rules for identifying troughs Forecasters had developed some practical rules for identifying conditions that implied the economy was near a trough. Haney’s (1931) forecasting textbook documents the judgmental non-statistical forecasting approaches commonly used during that period. He lists twelve conditions that should apply in the vicinity of cyclical troughs: 1. 2. 3. 4. 5. 6. 7. 8.

low interest rates; call-money interest rates below time-money rates; time-money interest rates below commercial paper rates; interest rates on short-time money below bond yields; production below normal; the ratio of inventories to production of manufactured goods is high but declining; unfilled orders are very low and have been declining; commodity prices were declining but are beginning to stabilize;

526

Articles

9.

the ratio of raw material prices to finished commodities prices is low but increasing; 10. employment is beginning to decline more slowly, or even to increase; 11. payrolls become stabilized, and 12. the stock market has turned up. The twelve indicators on this very comprehensive list can be divided into several categories. Indicators 1 to 4 and 12 refer to money market or financial market conditions. The level of economic activity in the goods and services sector is represented by conditions 5–7 and 10–11. Finally, the movement of prices is measured by conditions 8 and 9. Suppose one adopted this system of indicators for forecasting. The question then becomes: did the data for these indicators available near the end of 1930 justify the prediction that the economy was near a cyclical trough? Table 3 presents data available at the time relevant to these twelve indicators. These data do not consistently indicate that the economy is in the vicinity of a trough. On the one hand, the financial indicators largely conform to the postulated conditions, and forecasters continually refer to the easy money condition.13 On the other hand, the non-financial indicators do not indicate this. Production, factory employment and payrolls showed no improvement in any month in the third quarter of 1930. Commodity prices were still declining and the ratio of raw materials prices to finished goods prices reached new 1930 lows in both November and December. Finally, even the stock market was not yielding the expected condition, because it was still declining sharply during the latter part of 1930. Because of a declining stock market and falling prices, businessmen were pessimistic, but the forecasters said this was unwarranted. This analysis shows that the available data were not consistent with some of Haney’s practical rules for identifying a cyclical tough and did not justify a prediction of an imminent upturn in the economy. Both this example and our earlier discussion of inventory data show that the optimistic predictions were inconsistent with the available data. Can this disconnect between the predictions and the available data be explained by the prevailing forecasting methodologies? A more general overview of contemporary forecasting methods Hardy and Cox (1927: 15–19) indicate that the three most commonly used methods of business forecasting during that period were: historical comparisons, ‘cross cut analysis’, and ‘specific historical analogy’. Historical comparisons looked for the ‘rhythm of business activity’ in regularities in the ‘succession of economic events’. This implied that all cycles were similar and that one could forecast merely by examining the sequence of cyclical phases.14 However, Hardy and Cox note that each situation is different, and that it is necessary to identify the factors that

Table 3 Selected data for 1929–1930 Interest rates

Prices

Time Loans 7.75

Ratio: Call Treasury Industrial Mfg Indus- Commo- Material/ Loans Bonds ProductionProduction Total trial dity Finished 6.94 3.59 118 118 185 193 97 1.034

Factory Factory Dow-Jones Employ Payrolls Stock Prices 100 101 307.25

5.5 5.75–6 6 6 6 6 6–6.25 6.25 6.25 5.25–6.25 5 4.75–5

7.5–7.75 7.75–8 8.5–9 8.5–9 8–8.25 7.5–8 8.75–9 8.75–9 7.0–9.0 4.75–6 4.75–5 4.75–5

7.47 9.8 9.46 8.79 7.83 9.41 8.15 8.62 6.1 5.4 4.88 4.31

3.66 3.76 3.67 3.67 3.71 3.68 3.72 3.7 3.67 3.45 3.46 3.51

118 119 122 124 126 124 123 122 118 108 101 104

117 120 123 126 129 126 124 122 118 108 98 103

187 189 187 188 191 207 218 225 202 151 154 156

192 196 193 193 191 203 210 216 194 145 147 149

97 98 97 96 96 98 98 98 96 94 94 93

1.023 1.025 1.001 0.989 0.999 1.013 1.016 1.02 1.014 1.006 1.012 1.008

100 101 102 102 103 103 103 102 101 99 97 96

108 111 111 111 110 106 111 112 111 103 99 94

309 308.85 309.2 310.25 316.45 341.45 359.15 362.35 291.5 228.2 247.2 255.65

4.5–5 3.75–4.75 3.75–4 3.5–4 3.25–3.75 3–3.5 3 3 3 2.75–3 2.75–3

4.5–5 3.75–4.75 4–4.5 3.25–3.75 2.5–3.25 2.5–3 2.5–2.75 2.5–2.75 2.25–2.75 2–2.25 2–2.5

4.28 3.56 3.79 3.05 2.6 2.18 2.22 2.17 2 2 2.27

3.5 3.46 3.4 3.41 3.37 3.37 3.38 3.37 3.34 3.32 3.34

107 104 107 104 100 95 91 91 87 85 82

107 106 107 105 100 94 91 90 85 84 80

166 172 181 171 153 149 148 149 128 117 109

156 163 171 160 143 140 139 139 118 109 102

92 91 91 89 87 84 84 84 83 80 78

0.991 0.971 0.977 0.965 0.954 0.935 0.947 0.946 0.935 0.918 0.906

94 93 92 91 90 87 84 83 82 81 80

98 98 97 94 91 83 82 83 81 75 74

267.4 278.25 285.5 266.7 243.15 229.8 228.8 225 198.75 180.95 172.15

527

Comm Paper 5.25–5.5


Jan 1929 Feb Mar April May June July Aug Sept Oct Nov Dec Jan 1930 Feb Mar April May June July Aug Sept Oct Nov Dec

Inventories

528

Articles

produce economic changes. This is called ‘cross cut analysis’ and involves reasoning from causes to effects. Factors identified are classified as stimulative versus contractionary. A judgmental comparison and weighting of the positive and negative stimuli is used to produce the forecast. The use of historical analogies is, in essence, a combination of the other two methods. In this approach, previous cycles are examined to find one where there are similarities with the current conditions. If the ‘forces operating in the two instances are found to be similar, prediction is made on the basis of the events that followed the earlier situation’ (Hardy and Cox: 19). Of course, the analyst is expected to modify the forecast to take into account factors that may be somewhat different in the two situations.15 Another author (Haney 1931: 226–7) points out that the use of analogies is dangerous because all important factors must be taken into account and that may not be possible. Haney then argues that ‘analogies are probably most valuable in judging … when a turn may be expected’ (ibid: 227). This is precisely the way in which analogies were employed at the end of 1930. Persons’ (1931) use of analogies serves as an illustration of that forecasting method. Persons was one of the most prominent forecasters of that period.16 In November 1930 he predicted that the cyclical trough would occur between the end of that month and February 1931. His analysis was based primarily on reasoning by explicit analogy.17 He found that the situation in late 1930 was most comparable to the US depressions of 1884–5, 1907–8 and 1920–1. The maximum length of those depressions was 25 months and the maximum declines in activity were approximately 25 per cent. With the Index of Industrial Production already down by that magnitude, he concludes: ‘If the current depression is not to extend over a longer interval… recovery should begin in the next month or two…’ (Persons 1931: 17). He did consider whether there were factors, such as the international debt situation, installment buying and bank loans on securities, that might prevent a recovery but dismisses them. Moreover, he argues that there is a good reason why economic activity has not declined more than 25 per cent in the past and why it should not exceed that amount in this depression. ‘Drastic business recession ultimately encounters the resistance of the purchases of consumers who utilize their resources to protect their standards of living’ (Persons 1931: 23). The Hardy and Cox forecast classification scheme discussed above implies that Persons was not alone in using analogies to forecast. As some explicit evidence of the frequent use of analogies, we provide in Appendix A3 instances of explicit analogies from the forecasts we tabulated earlier in this paper. To cite two examples from that Appendix, Detroit’s Union Guardian Trust asserted in early August 1930 that improvement in economic conditions should be expected in that month because recessions usually last 12–15 months. Chicago’s Foreman State National Bank indicated in October 1930 that the 1921 upswing began 12 months after the previous


529

peak. Since there had been an upswing in 1930, the next upswing would begin 10–12 months from then. Why were the analogy-based forecasts so wrong? Clearly, Persons’ analogy-based forecast was seriously flawed, as were those cited in Appendix A3. However, there is no well-established methodology for evaluating forecasts primarily based on analogy. Nonetheless, one can examine whether the method was used appropriately given the knowledge that was available in 1930. This provides an evaluation based on ex ante or real-time reasoning. The next step is to note the factors that were overlooked or could not have been predicted at the time. This step provides material for an ex post evaluation of these analogy forecasts. We have already noted that the use of analogies was a very wellestablished method of forecasting in the early 1930s. Moreover, an examination of Persons’ (and other forecasters’) discussions indicates that attention was paid to factors that might make 1930 different from previous cycles. Thus, we conclude that, in real time, the forecasters of that period used this method of forecasting in an appropriate manner. However, an ex post evaluation can help explain why the forecasts were so wrong. Ex post evaluation of the 1930s analogy forecasts An analogy is a comparison between two ‘similar’ situations or objects. To be able to draw useful ‘comparison-inferences’ from the analogy, the two things being compared must be ‘similar enough’ in ways essential to the comparison at hand. There are at least two important aspects of this notion of ‘similar enough’. First, the two situations or objects might have many characteristics in common, making them seem ‘virtually indistinguishable’. Nevertheless, they might differ in a single characteristic that is so crucial for the analogy application in question that in essence they are not usefully comparable. Thus, forecasting-by-analogy may fail even if there is only one characteristic that differentiates the two objects or situations, but that one characteristic is crucial to the way that the economy functions.18 Second, and in opposition to the previous point, there are examples of useful analogies where characteristics that might be thought important fail to match, yet the analogy is useful nonetheless. An interesting example is that in a particular context water provides an excellent analogy for electricity in the following sense. An understanding of the relationship embodied in Ohm’s Law between resistance, current and voltage can be developed through a hydraulic analogy.19 One important use of analogies in forecasting is to make inferences from the ‘observed’ features of one of the two allegedly-comparable items or situations to an ‘unobserved’ feature of the other allegedly-comparable item

530

Articles

or situation. For this kind of analogy use, one must first decide whether the two situations or items or events are similar enough with respect to ‘forecast-essential’ observable features. If the two are similar enough in these essential features, then one extrapolates from a feature of interest that is known for one of the entities to that same unobserved feature of the other. How might this ‘is it analogous enough?’ concept apply to business cycle forecasting in general, and the 1930 forecasting situation in particular? Suppose that we are in the midst of a recession and want to predict when it will end. We know how long previous recessions have lasted and want to use this ‘known’ information to estimate by analogy how long the current one will last. Is this recession ‘similar enough in all essential observable characteristics’ to one or more of the past recessions? If it is, then ‘forecasting by analogy’ may seem warranted. That is, it may seem plausible to estimate the length of the current recession based on the observed length of the comparable previous one(s).20 Let us see how this criterion applies to the analogical reasoning Persons used to produce his 1930 forecast-by-analogy. Persons had considerable information about all previous recessions, including their durations. He also had a substantial amount of data about the 1929–30 recession, but not how long it would actually last. If other crucial characteristics of some earlier cycles were ‘similar enough’ to 1929–30, then a forecast of the duration of the 1929–30 downturn could plausibly be based on the duration of the earlier comparable downturns. This is exactly the kind of analogical reasoning Persons used. As we noted above, Persons’ argued that the current depression has the ‘closest parallels’ to those of 1884–6, 1907–9 and 1920–1. He then went on to argue that ‘The parallelism of 1907–08, and 1920–21 with 1929–30 is less perfect than … 1884–85’(ibid: 14).21 Persons draws out the similarities between 1884–5 and 1929–30, stressing that all the leading commercial countries were depressed, that there were rapidly falling commodity prices, that imports to the US were low relative to prior years, that construction (railroads in the 1880s, new buildings in 1929– 30) had (after booming) declined severely, that capital was flowing out of the country, whereas previously it had been flowing in, that stock prices were declining, that banks and brokerages were failing, and there was a brief money panic. The implication for Persons was that, whether one restricted oneself to the 1884 analogy, or also considers the other two periods as sufficiently analogous, one gets ‘substantially consistent pictures of the ‘pattern’ of a major business depression and recovery in the United States’ (ibid: 16). The pattern involves a ‘maximum interval of subnormal business’ of 25 months, a maximum trough length of 7 months and ‘intervals of recovery to ‘normal’ … from 12 to 14 months’ (ibid: 16). The implication of the analogy is that ‘if the current depression is not to extend over a longer interval than those of


531

1884–85, 1908 and 1921 … recovery should begin in the next month or two and normal business should be attained by the end of 1931’ (ibid: 17). 22 So what flaws in this analogy might have led to its failure? One possibility is that the list of characteristics Persons uses that make the 1929–30 depression seem analogous to the three earlier recessions may be seriously incomplete. For example, Persons was aware of the deflationary pressures, but he did not explicitly consider the relationship between indebtedness and deflation. Irving Fisher (1933) argues that the effect of falling commodity prices in the presence of large-scale indebtedness was crucial in extending the 1929 downturn.23 In Fisher’s analysis, the prior recessions most similar to 1929–30 are 1837, 1873 and 1884, all of which had this deflation-and-debtoverhang characteristic. If Fisher is right, Persons made an error in his inferences from these cases.24 An economic historian could probably find other characteristics that differed in the cycles. Implications for the use of analogies in forecasting This analysis of the analogies used in 1930 illustrates the kind of mistake that can be made when forecasts are based on explicit analogies.25 However, it is simply incorrect to think that there are some standard forecasting techniques that are guaranteed to avoid the potential pitfalls of forecasting-by-analogy. The reason is that virtually all serious forecasts contain analogy elements, although these analogy elements are often implicit rather than explicit. Virtually all serious forecasts involve (sometimes implicit) analogy elements because they are based on prior relationships, which are then used to make inferences about future events based on current knowledge. This is implicit analogical reasoning, resting on the assumption that the future is likely to be ‘similar enough’ to the past to generate a legitimate inference about the future from known past experience and current conditions. Thus, a GDP forecast based on a statistical relation such as a regression indicates how the level of economic activity is related to a set of monetary and/or real variables. This relationship has been estimated from past data. Structural models do this in a more analytical way than reduced form models,26 but both implicitly assume that some identified relationship (in the structural estimation case, it is the estimated underlying ‘structure’ of the economy) will continue to hold in the future: the future relationship will be analogous to the past.27 Consequently, when new data are inserted into the estimating equation(s), it is expected that the model will yield an accurate prediction.28 There are at least two implicit assumptions in this approach. First, it assumes that variables that have been omitted from the estimating equation(s), because they did not have an important effect in the sample period, will also not matter beyond the period of fit. Second, it is assumed that there is no structural change so that the basic relationships continue to hold. If these assumptions are violated, the quantitative method can

532

Articles

generate the same types of forecast errors that arise from explicit forecasting-by- analogy.29 A useful example As a general matter, forecasters fail to predict recessions prior to their occurrence and rarely predict one that does not occur (Fildes and Stekler 2002). Forecasts made for 1979 provide one interesting exception, since a recession was expected to occur that year, but it did not materialize. It is possible to explain why this false prediction occurred. There was an oil shock in 1973 and it produced a recession. There was another oil shock in 1978, and using the analogy with 1973, economists predicted a recession. In hindsight it is clear that there was a crucial difference between the two situations. The Economic Report of the President (1980: 30ff) explains why the economy was more resilient than had been expected. First, consumers had experienced inflation and had learned how to adjust their consumption purchases when inflation accelerated. Their expectations of inflation were that an increase in the inflation rate would be followed by an additional increase in inflation. This caused them to accelerate purchases in anticipation of even higher prices in the future. Whereas the savings rate had increased in 1973, it declined in 1979. A second explanation for the resiliency of the economy was the financial deregulation that had occurred. As a result monetary policy worked more through the interest rate channel and less through credit availability. As a consequence, monetary restraint no longer produced an abrupt change in credit availability. Finally, there were no cyclical imbalances such as an inventory overhang. In terms of our earlier discussion of forecasting by analogy, there was an important characteristic, the oil shock, that was similar in both 1973 and 1978. However, other factors such as the inflationary expectations and the financial structure differed between 1973 and 1978. It is these dissimilarities that made the implicit ‘forecast-by-analogy’ inappropriate and therefore incorrect. CONCLUSIONS A contribution of this paper has been to document that many forecasters produced overly optimistic forecasts in late 1930, predicting that the trough would be reached and recovery begun long before it actually happened. We then examined two important issues: (1) why the 1930 forecasts were so wrong and (2) what implications these underlying errors have for modern forecasting methodology. Our analysis of the 1930 forecasts indicated that an important element of the economic theory available at the time was recognized in the preparation


533

of the forecasts. Moreover, the forecasting methodology for identifying turning points was well established at that time. We also demonstrated that the real-time data did not correspond with either the theoretical criteria or the forecasting methodology for the existence of a trough. Nevertheless, the forecasters continued to produce optimistic predictions. This analysis has implications for current forecasting methodology. However, a caveat is necessary before setting out these implications, because our analysis focused on a very particular type of sample. We have not taken a random sample of all prior forecasts and studied that sample for the incidence of errors, and for the factors seemingly associated with such errors or with the incidence of forecast successes. Such an analysis would, for example, enable us to determine whether forecast successes or failures were due to chance. Instead, we have taken a very special set of ‘bad forecasts’ and tried to draw inferences from that special sample about the sources of error. On the other hand, our procedure of trying to learn from what seem like clear mistakes is similar to the fairly common methodological practice in several settings: US hospitals, the National Transportation Safety Board (NTSB) and NASA. In the hospital setting, staff doctors meet to discuss treatment cases where something seems to have gone badly wrong, to try to identify the sources of error. An important aim is to reduce the probability of future recurrence of these errors. The NTSB ferrets out the causes(s) of specific aircraft crashes, and NASA used similar procedures in investigating the Challenger and Columbia disasters. Presumably, one motivation is to make improvements in order to prevent future disasters from ‘related’ causes. So why did the 1930 errors occur, and what can forecasters learn from these errors? One referee suggested that the errors were due to ‘sensible optimism’, meaning that forecasters ‘like others at the time, might have wanted the bad times to go away and so they bent their forecasts to their desires’. This is a form of bias that can recur if the same forecasting process is used. This optimism might not have occurred if the 1930 forecasters had questioned why there was a disconnection between the data (including the inventory/production data), and Haney’s twelve conditions (especially declining prices) on the one hand, and the theory and predictive processes on the other. Consequently, forecasters might learn from this experience to always question whether the available data show that the economy is on the track that is implied by the forecast.30 An alternative explanation for the 1930 forecast errors is that they were ‘honest mistakes’ resulting from the forecasting methodology that was available at that time. One example was that the forecasts in question were sometimes based on inappropriate analogies with previous recessions; forecasters failed to appreciate the unusual non-analogous features of the Depression. This finding has implications for current forecasting methodology because current practice contains implicit analogy elements. Forecasters can ‘learn’ from this analysis to always check whether the implicit

534

Articles

assumptions inherent in the forecast are likely to still hold or, to the contrary, structural change or an unusual situation may have occurred. There is an instructive comparison to be made here with recent work on the methodology of modeling. Morgan and Morrison (1999) suggest that models themselves sometimes incorporate analogical elements. The suggestion itself reflects earlier work in the philosophy of science (see, for example, Hesse 1966). Our reading of Morgan and Morrison is that these analogical elements have two aspects: models as representations of (analogies to) reality, and models in comparison to (features analogous or disanalogous to) other models. The analogy element in forecasting has a somewhat different focus, involving the assumption that the relevant ‘world’ has not changed in crucial ways: the structure of the future is analogous to the structure of the past. Morgan’s concluding chapter in Morgan and Morrison (1999) is entitled ‘Learning from Models’. Since the analogical element in forecasting is not identical to the analogical elements in modeling, one can ask how ‘learning from forecasts’ might take place. The need for the past-future analogy to hold if forecasts are to be valid suggests a hypothesis. Regardless of the particular explanations for those 1930 forecast errors, one hypothesis might be: ‘Forecasters learn things one error at a time’.31 Robert S. Goldfarb George Washington University [email protected] H. O. Stekler and Joel David [email protected] NERA [email protected]

ACKNOWLEDGEMENTS The authors thank Ed Gamber, Kevin Hoover, Fred Joutz, David Ribar and two anonymous referees for many helpful comments on an earlier draft. We are responsible for any remaining errors.

NOTES 1 2

Hamilton finds in addition that, while deflation was expected to occur in the latter part of that year, the actual decline in prices was more pronounced. Even though businessmen expected deflation to accelerate in the latter part of 1930, they were also forecasting a recovery in economic activity (Nelson 1991: 37). The forecasts of the Harvard Economic Society said that prices would not rise until activity went up.

1930 forecasting errors 3

4

5

6 7 8 9

10

11 12

13 14

535

Although the 1930 meetings of the American Economic Association had a session entitled ‘The Business Depression of Nineteen Hundred Thirty’, most of the discussion was about the causes of the depression, its severity and difficulties of recovery. There was no discussion in the published proceedings of when the depression would end. Thus, we do not have forecasts by economists from those meetings. These analyses are issued monthly and have the same time frame as the monthly forecasts. Although there are also weekly indexes (and commentaries) that measure economic activity, they do not have the same time frame and, moreover, can be very volatile, depending upon the events that occurred in that week. There was no need to reconcile the scores that the two authors independently assigned to the qualitative assessments published in the Federal Reserve Bulletin because the two sets of scores were virtually identical in every month. The assessments published in the Bulletin were more precise and less ambiguous than the business forecasts. This may explain why there was more agreement in the authors’ independent scores of the FED’s statements. Appendix A4 displays the scores for each individual forecaster. Smith (1931), using regression analysis with leading indicators as the regressors, also failed to predict what would happen in 1931. This quote comes from an article, written in November–December 1930 that was reprinted as Chapter 1 of Persons’ 1931 book. Our analysis compares the forecasts (made at time t) with the ‘actuals’ as measured by the FED’s description of the economy averaged over the three months t, t+1, and t+2. The FED’s analysis covers all sectors of the economy (production, employment, distribution), as well as the financial sector and commodity prices. Friedman and Schwartz (1963) and Bullock and Crum (1932) indicate that there was a minor recovery in early 1931. If this were true, our evaluation of the forecasts would not be as negative. However, our scoring of the FED’s descriptions was not as positive as the comments of Friedman and Schwartz or Bullock and Crum. The FED recognized the upturn in production that occurred but also noted the weaknesses in the other sectors including the continuing deflationary pressures. Most articles about forecasting appeared in The Journal of the American Statistical Association. However, Marget’s (1929) review of Oskar Morgenstern’s book, a tome containing the assertion that economic forecasting was impossible, appeared in the Journal of Political Economy. For a contemporary explanation that this was one condition for economic revival, see Haney (1931: 159). Haberler (1937), writing after the Depression, enumerates many additional factors that can create recovery. All data are seasonally adjusted. The inventory data published in the 1931 and 1932 Annual Supplements to the Survey of Current Business were not seasonally adjusted. Professor Fred Joutz seasonally adjusted these data for us using the EVIEWs program. It should be noted that Haney (1931: 165) also graphed this relationship and that his graph is similar to ours. The condition (1) for low interest rates does not distinguish between real and nominal rates. Although Hardy and Cox do not reference the research of the National Bureau of Economic Research, Mitchell (1927) and others were at this time analyzing the various phases of the business cycle. A thorough and illuminating discussion of business cycle research, including the work of Mitchell, in the context of the development of econometric ideas is Morgan (1990), especially Chapters 1 and 2. Judy Klein cites Mitchell’s work in the

536

15

16

17

18 19 20

Articles context of what she calls the ‘relative time framework’ (Klein 1997: 104ff and 298–9; Mitchell is discussed on 134). This framework involves treating ‘every data point as a place in a sequence’ (104). The Mitchell/NBER view of business cycles as having ‘stages’ fits this ‘relative time’ idea (Klein: 134). A similar classification scheme is presented by Edmund Day in his 1927 presidential address to the American Statistical Association, published in the March 1928 issue of JASA. After asserting that business forecasting and the discipline of statistics ‘are assuredly joined for life’, Day suggests that there are ‘three general types of prediction:(1) prediction by analogy; (2) prediction by formula; (3) prediction by analysis’ (1928: 2). ‘Prediction by formula’ turns out to be very similar to Hardy and Cox’s historical comparisons, and ‘prediction by analysis’ very similar to Hardy and Cox’s ‘cross cut analysis’. A striking difference is in the two sets of authors’ view of prediction by analogy. Hardy and Cox view analogy as a combination of the other two categories, while Day views it is the simplest and most naive (and non-statistical) of the three. Morgan (1990: 40) notes that ‘Mitchell and Persons’ … empirical programmes dominated statistical business cycle research in the 1920s and 1930s’. One illustration of Persons’ stature among his contemporaries is in van den Bogaard (1999). He argues that the development of the Dutch ‘barometer’ method of dealing with time series data, ‘and of other European barometers as well, was heavily influenced by the work done by the Committee on Economic Research which had been appointed by Harvard University in 1917…. Many aspects of the method used and the accompanying arguments seem to be almost literally copied from the work of W. M. Persons, who worked for this Committee …’(van den Bogaard 1999: 288). Morgan (1990: 56–63) discusses two major articles Persons published in 1919, characterizing some of the work as ‘primarily responsible for developing data preparation and data adjustment methods which quickly became and remained standard in applied econometrics’ (ibid: 57). She notes that Persons resigned from the Harvard faculty in 1928 to become an economic consultant (ibid: 57, footnote 12). Persons uses the term ‘analogy’ explicitly in his analysis: for example, ‘the analogy between the economic backgrounds of 1930 and 1884–85 is closer than …’ (1931: 14). A related term he relies on heavily is ‘comparable to’: ‘the severity of the present depression is comparable to the major depressions of the last 55 years …’ (ibid: 7–8). Moreover, he repeatedly makes arguments like the following: ‘the first step to securing a notion of the probable future developments of industry and trade is to ascertain how industry and trade have actually fluctuated over a long period of years in the past’ (ibid: 33). Note that his 1931 book Forecasting Business Cycles reproduces some material written earlier, such as his November 1930 forecast. For further discussion of the nature of analogies and their use in public policy analysis, see Goldfarb and Henig 2000. We owe this point about the water-electricity analogy to comments on an earlier draft from Kevin Hoover, whose description of the analogy we have paraphrased. As noted above, Haney (1931), indicates that ‘comparisons with the past … may be of value …. [but] are dangerous’. He then suggests that analogies ‘are probably most valuable in judging about when a turn may be expected’ (Haney 1931: 226–7). However, he then provides the following warning under the heading ‘DON’TS FOR THE FORECASTER’: ‘Don’t base forecasts on mere analogies with past cycles. No two are exactly alike, and allowance must


21

22 23 24

25

26

27

28 29 30 31

537

be made for different causes and different conditions’ (ibid: 230). Mitchell (1927: 354) had also noted that every cycle was different. He argues for example, that ‘in 1907–08, although a moderate cyclical decline of commodity prices occurred, this decline was merely an episode in a longer period of advancing commodity prices; and in 1921 the existence of postwar shortages in building and certain manufactured goods … stimulated recovery’ (Persons 1931: 14). Klug et al. (2002: 4) also suggest that railroad shippers used an analogy in forecasting: ‘like the professional forecasters … railroad managers failed to forecast the depression, expecting it to follow the pattern of recent recessions’. For a related argument, see Romer (1993: 25–6). Persons did use 1884 as analogous, but did not discuss the combination of huge indebtedness and deflation. More generally, by the end of 1930, the forecasters of that period were aware of both the deflation that was occurring and the high level of consumer installment debt. Yet they never connected the two to realize that their interaction would impose a huge burden on consumers. King (1932) reports on a conference that conducted a post mortem analysis of forecasting methods used during the Depression. One of the conclusions was that empirical methods are unsound if they depend upon the number of times that a certain relationship has recurred in the past. While the analogies used in 1930 were not explicitly mentioned, they would fit into this category. Both techniques assume that the future is like the past, but in the case of structural models, the attempt is to capture the economy’s underlying structure. If this is done, then even if ‘surface events’ of the future are different from those of the past, the forecasts, based as they are on the underlying real structure, might still be appropriate. Recent work in the philosophy of science about the nature of models, including models in economics, indicate that models themselves sometimes incorporate analogical elements. See Morgan and Morrison (1999) for a discussion, including cites to the philosophy of science and economic methodology literatures (see especially pp. 5–17), and applications to particular sciences, and to economics (see especially ch. 12). It is well known, of course, that the variance of the forecast error will be larger than that of the estimating equation. An anonymous referee, reacting to this point that forecasts typically have implicit analogy elements, suggested that this is just another instance of ‘our old friend the induction problem’. Stekler (2003) suggested that this type of questioning is necessary to identify ‘unusual’ events, and the Great Depression certainly was in that category. We owe this phrase to Ed Gamber.

REFERENCES Bullock, C.J. and Crum, W.L. (1932) ‘The Harvard Index of Economic Conditions: interpretation and performance, 1919–31’, The Review of Economic Statistics 14: 132–48. Cecchetti, S.C. (1992) ‘Prices during the Great Depression: was the deflation of 1930–1932 really unanticipated?’, American Economic Review 82: 141–55. Day, E. (1928) ‘The role of statistics in business forecasting’, Journal of the American Statistical Association 23(161) (March): 1–9. Dominguez, K.M., Fair, R.C. and Shapiro, M.D. (1988) ‘Forecasting the Depression: Harvard versus Yale’, American Economic Review 78: 595–612.

538

Articles

Evans, M. and Wachtel, P. (1993) ‘Were price changes during the Great Depression anticipated?’, Journal of Monetary Economics 32: 3–34. Fildes, R. and Stekler, H. (2002) ‘The state of macroeconomic forecasting’, Journal of Macroeconomics 24: 435–68. Fisher, I. (1933) ‘The debt-deflation theory of Great Depressions’, Econometrica 1: 337–57. Friedman, M. and Schwartz, A. (1963) The Great Contraction 1929–1933, Princeton: Princeton University Press for the NBER. Goldfarb, R. and Henig, J. (2000) ‘Coping with information problems in public policy analysis: the analogy ‘therapy,’ mimeo. Haberler, G. (1937) Prosperity and Depression, Geneva: League of Nations. Hamilton, J.D. (1992) ‘Was the deflation during the Great Depression anticipated? Evidence from the commodity futures market’, American Economic Review 82: 157–78. Haney, L.H. (1931) Business Forecasting, New York: Ginn and Company. Hardy, C.O. and Cox, C.-V. (1927) Forecasting Business Conditions, New York: The Macmillan Co. Hesse, M. (1966) Models and Analogies in Science, Notre Dame: University of Indiana Press. King, W.I. (1932) ‘Forecasting methods successfully used since 1928’, Journal of the American Statistical Association 27: 315–19. Klein, J. (1997) Statistical Visions in Time: A History of Time Series Analysis, 1662– 1938, Cambridge: Cambridge University Press. Klug, A., Landon-Lane, J.S. and White, E.N. (2002) ‘How could anyone have been so wrong? forecasting the Great Depression with the railroads’, National Bureau of Economic Research Working Paper 9011, mimeo. Marget, A.W. (1929) ‘Morgenstern on the methodology of economic forecasting’, Journal of Political Economy 37: 312–39. Mitchell, W.C. (1927) Business Cycles—The Problem and its Setting, New York: National Bureau of Economic Research. Morgan, M. (1990) The History of Econometric Ideas, Cambridge: Cambridge University Press. Morgan, M. and Morrison, M. (eds) (1999) Models as Mediators, Cambridge: Cambridge University Press. Nelson, D.B. (1991) ‘Was the deflation of 1929–30 anticipated? The monetary regime as viewed by the business press’, Research in Economic History 13: 1–65. Persons, W.M. (1931) Forecasting Business Cycles, New York: John Wiley and Sons, Inc. Romer, C.D. (1990) ‘The Great Crash and the onset of the Great Depression’, Quarterly Journal of Economics 105: 597–624. Romer, C.D. (1993) ‘The nation in Depression’, Journal of Economic Perspectives 7: 19–39. Smith, B.B. (1931) ‘A forecasting index for business’, Journal of the American Statistical Association 26: 115–27. Stekler, H.O. (2003) ‘Improving our ability to predict the unusual event’, International Journal of Forecasting 19: 161–63. van den Bogaard, A. (1999) ‘Past measurements and future predictions’, in M. Morgan and M. Morrison (eds) (1999) Models as Mediators, Cambridge: Cambridge University Press, pp. 282–325.


539

APPENDIX A1 FORECASTERS WHOSE FORECASTS ARE INCLUDED Table A1 Babson

Cleveland Trust

Foreman-State National Bank, Chicago

Northwest Bank Corporation, Minneapolis

Bank of America

Continental Illinois

Guaranty Trust, New York

Ohio Saving Bank & Trust, Toledo

Bank of Manhattan

Cox, Prof. Garfield

Haney, Dr. Lewis

Security National Bank, Los Angeles

Brookmire

First National Bank Chicago

Harvard Economic Society

Silberling Research Corporation

Central Trust, Illinois

First National Bank, Detroit

Hibernia National Union Guardian Bank, New Orleans Trust, Detroit

Chase National Bank

First Union Trust, Chicago

Irving Trust

Union Trust, Cleveland

National City Bank, New York

Wells Fargo

Chatham Phenix First Wisconsin National Bank National Bank Chemical Bank

Fletcher American National Industrial Wisconsin Bank Conference Board Shares Corporation, National Bank, Milwaukee Indianapolis

APPENDIX A2 SCALING SYSTEM FOR QUANTIFYING QUALITATIVE DESCRIPTIONS OF CURRENT CONDITIONS AND FORECASTS Table A2 General mindset Optimism

Condition diagnosed or forecast Return to ‘normal’ soon Strong vigorous recovery Recovery soon, definite signs of bottom Slight increase, smaller reduction than usual

Score assigned +1 +3/4 +1/2 +1/4

Neutral

Offsetting influences, irregularities; stabilized; usual seasonal increases

0

Pessimism

Negative direction with some offsetting factors Decreases more than usual, further overall declines Substantial, sharp decreases, considerable declines in activity Exceptionally rapid declines of large magnitudes

21/4 21/2 23/4 21

540

Articles

APPENDIX A3 PARAPHRASES OF ANALOGIES IN 1930 FORECASTS QUOTED IN THE COMMERCIAL AND FINANCIAL CHRONICLE Table A3 4/26/30, p. 2869

6/7/30, p. 3951 8/2/30, p.700 10/4/30, p.2133 10/18/30, p. 2419 10/18/30, p.2444 10/25/30, p.2598

11/1/30, p.2801 11/3/1930

11/15/30, p.3110 12/6/30, p.3613 12/20/30, p.3943

Chase

In comparison with 1921, this is mild. There is no crisis. Those come when there is an overextension of credit—excessive inventories financed by borrowing or long-term investment financed by short-term money. National City Recovery always occurs before the Bank second year of a recession Union Guardian Improvement [expected in] August Trust Detroit because recessions usually last 12–15 months National City Given that indexes are at 1921 levels, Bank we must be near the bottom Editorial the The decline has been so deep and so Chronicle long that [recovery] has to be soon Cleveland The long length of the recession makes Trust one believe that recovery will develop soon Foreman State In 1921, the upswing began 12 months Natl Bank after the peak. There was an upswing Chicago in 1930, so the upswing will begin 10–12 months from then Silberling 1929–30 is as far below normal as Research can be expected National Considering the months of reaction City Bank behind us, as well as the depths to which the decline has gone, there can be little doubt that the recession is scraping bottom. (Quoted in New York Times, p.35) Chase Comparisons with 1921 fundamentally do not justify such a low level of activity. Inventories were bad in 1921. National City Bank The Depression is 16 months old; some sectors have fallen 35%. The assumption is that the bottom is near. Cleveland Trust Based on past experience, no such depression has declined so much and so long without recovering

Cox, Prof. Garfield

Haney, Dr. Lewis

New York

Guaranty Trust,

Bank, Chicago

Foreman-State National

Bank, Indianapolis

National

Fletcher American

National Bank

First Wisconsin

Chicago

First Union Trust,

Detroit

First National Bank,

Chicago

First National Bank

2.25, 2.25 2.5, 2.75

2.5, 2.25

2.5, 2.25

2.5, .5

na, 0

2.25, na

0, .5

2.5, 2.5

Cleveland Trust

Continental Illinois

2.25, .25

Chemical Bank

National Bank

Chatham Phenix

na, 2.5

0, na

1/30

na, na

2.5, na

2.25, 2.5

12/29

Chase National Bank

0, 2.25

0, 2.25

11/29

2.25, 2.5

2.5, na

10/29

Central Trust, Illinois

Brookmire

Bank of Manhattan

Bank of America

Babson

Name

Table A4

.5, 2.2

0,.5

2/30

.5, .5

2.25, na

21, 2.25

.5, .5

.25, 0

3/30

0, .25

2.25, .25

0, .25

0, na

0, 0

4/30

0, 2.25

2.5, .25

2.5, .25

.5, .5

.25,na

5/30

0, 0

.25, na

0, 0

na, na

6/30

2.5, 0

2.25, 2.5

2.25, .25

.25, na

2.25, .25

.25, .25

2.5, .5

7/30

2.75, .25

20.5, .25

na, 25

8/30

2.25, .25

0, .25

na, .25

10/30

2.5, 2.25 2.25, 2.25

.25, 0

0, .25

na, .5

9/30

2.25, .25

2.25, 0

2.5, na

0, na

2.5, na

0, .25

11/30

0, .25

.25, .25

0, .25

2.75, na

0, .25

12/30

1930 forecasting errors 541

APPENDIX A4

FORECASTS OF INDIVIDUAL FORECASTERS

542

Name

10/29

11/29

12/29

1/30

2.5, 2.25

2.5, 2.75

2.75, 2.25

2.25, .25

Harvard Economic Society

2/30

3/30

0, .25 2.25, .25

4/30

5/30

6/30

7/30

8/30

9/30

10/30

11/30

12/30

0, .5

0, .5

0, .5

2.5, .25

2.75, 0

2.5, .25

2.25, 2.25

2.75, 2.25

2.5, .25

.25, na

.25, na

0, na

2.25, 2.25

2.75, na

0, 0

2.25, 2.25

0, .25

0, .25

2.25, .25

0, 0

.25, .25

2.5, .5

2.75, .25

2.25, 2.25

0, na

2.25, na

2.75, .25

2.5, na

.25, na

.25, .25

2.25, .25

Hibernia National 2.5, 2.5

Bank, New Orleans Irving Trust National City Bank, New York National Industrial Conference Board

na, .25

0, .25

225, 2.25

0, na

2.25, na

2.5, na

na, na

2.75, .5

2.5, na

.25, na

0, .25

2.25, na .25, .5

2.75, .25

2.5, .25

Northwest Bank Corporation,

0, na

Minneapolis Ohio Saving Bank &

2.25,

Trust, Toledo

.25

Security National Bank, Los Angeles Silberling Research

21, 21

Corporation

Union Trust, Cleveland

21,21

0,na

21, 2.5

2.75,2.75

0, na

0, na

.25,2.75

0, .25

0, .25

2.5, 2.5 2.5,2.75

Union Guardian Trust, Detroit

21, 21

2.5,na

2.25, 2.25

na, .75

0, 2.5

2.5, na

0, na

2.25,.25

0, .25

2.5,

Corporation, Milwaukee

.25, .25

2.5, .5

0,0 2.25, 0

0, .5

na, .5

na, .25

0, .25

2.25, .25

.25, na

.25, .25

0, 2.25

0, .5

2.25, .25

2.5, .25

2.75 2.5, .75

0, .5

.25, na

.5,.25

0, na

2.25, .25

0, .5

2.5, .25

Wells Fargo Wisconsin Bank Shares

2.5, .5 21, 0

2.75, na .25, .5

2.75, .5

0, na 2.25, .5

2.25,.25 2.25, .25 na, na 2.25, .25

0, .5

Articles

Table A4 (continued?)