Technology Forecasting

Technology forecasting is three activities wearing one name, and almost every argument about whether it works is an argument about which one is meant. Rate forecasting asks how fast the cost or performance of an existing technology will improve, and extrapolates a measured trend. Date forecasting asks when a capability will exist, and needs a threshold definition before it can be scored at all. Diffusion forecasting asks how far and how fast something already built will spread, and fits a saturating curve.

The framing under test is that technology can be forecast and the methods are known. The methods are known and separately validated; the framing still fails. The class with the good evidence base — improvement rates, hindcast over thousands of forecasts — is not the class anyone asks about. The class everyone asks about has one large scored corpus and it returns a third of forecasts inside a generous timing window. And the single most famous success in the field turns out, on the testimony of the industry that produced it, not to be a forecast at all.

Established Start with the strongest result, because the field's reputation is set by its weakest. The experience curve begins with T. P. Wright, then of Curtiss-Wright, in the Journal of the Aeronautical Sciences for February 1936: a power law in cumulative output, an exponent of −0.322, glossed as an “eighty percent curve” — double the quantity, and average labour cost falls to 80% of what it was. Established The empirical base was two or three points. Wright says so: the curve “was worked up empirically from the two or three points which previous production experience of the same model in differing quantities made possible.” No systematic set of models, no named production runs, one firm's cost records. The foundational paper of the entire experience-curve literature is an admitted handful of observations.

Established The relationship survived the thinness of its origin, and the comparative test is the reason to believe it. Nagy, Farmer, Bui and Trancik put five candidate laws head to head on 62 technologies across chemicals, hardware and energy, each with 10 to 39 annual observations from the Santa Fe performance-curve database. Wright's law — cost falls with cumulative production — forecast best. Moore's law — cost falls with time — came second. Goddard, the Sinclair-Klepper-Cohen hybrid and a lagged Wright all did worse; Nordhaus was excluded for overfitting. Overall forecasting error grows at roughly 2.5% per year of horizon. Frontier The paper's own explanation cuts against the field's self-description. Production itself tends to grow exponentially, so the two laws are nearly collinear. If the mechanism-free model fits almost as well as the mechanism, the fit does not identify learning-by-doing as the cause, and every policy argument that treats deployment subsidy as a lever on cost is resting on an identification nobody has made.

Established The best-validated claim in technology forecasting is a claim about intervals, not about points. Farmer and Lafond model technology cost as a correlated geometric random walk with drift and derive a closed form for the distribution of forecast errors as a function of horizon. 53 technologies with statistically significant improvement rates, filtered from 66. The validation is exhaustive hindcasting: with a five-year historical window there are 8,212 possible forecasts, of which 6,391 fall at horizons of twenty years or less. Normalised mean squared error grows as the square root of the horizon; with an autocorrelation parameter of 0.63, errors across all technologies and all horizons collapse onto a single Student-t distribution that survives surrogate-data testing. Their worked example is characteristically shaped: “the probability is roughly 5% that in 2030 the price of solar PV modules will be greater than or equal to their current (2013) price.” Established The method does not say what a cost will be. It says how wrong you should expect to be, and that estimate has been checked out of sample thousands of times. Quoting the central path without the interval discards the entire contribution.

Established Against that, the profession's actual record on the flagship technology of the last two decades. Way, Ives, Mealy and Farmer collected 2,905 projections by integrated assessment models of the annual rate at which solar photovoltaic system investment costs would fall between 2010 and 2020. The mean projected rate was 2.6% a year. Not one exceeded 6%. Costs fell 15% a year. Over a longer window solar, wind and battery costs have dropped roughly exponentially near 10% a year, and solar module cost has fallen more than three orders of magnitude since first commercial use in 1958. Established The shape of the error is the finding, not its size. If forecast error were noise, roughly a third of 2,905 independent professional projections would have been too optimistic. None were. Errors that all point the same way are not a noisy measurement of the future; they are an accurate measurement of the modelling community's priors. Frontier Mark the interested party: Way et al. are the same research group as Farmer and Lafond, and the paper both criticises other people's forecasts and advertises the authors' method. The 2,905 count is a count of other people's published numbers and is unaffected; the paper's headline savings estimates — about $14 trillion expected net present value for a fast transition at the Stern Review's 1.4% discount rate, median about $26 trillion — are attributed here, not asserted.

Established Now the other class, where the evidence is bleak and the corpus is small. The largest scored body of dated technology forecasts located anywhere is a 2012 contractor report by Carie Mullins for The Tauri Group, filed with the US defence technical information centre. 2,092 assessable forecasts extracted from 300 documents; 1,055 verified and scored. Results: about 70% — 736 — of predicted events had happened at all. Only 33% succeeded when the forecast also had to land inside a timing window of plus or minus 30%. By method, quantitative trend analysis scored 45% and models 37%, with the report concluding that “forecasts generated from quantitative trend analyses have statistically higher success rates than do forecasts generated from other methodologies.” By area, computer technology led at 40%. By horizon: short-term 35%, medium 28%, long 27%. The scored bias is that “forecasts are more likely to overestimate the event date” — a reversal of the earlier 310-forecast study, which found no systematic direction at all.

Established The gap between 70% and 33% is the whole of date forecasting in two numbers. Predicting that something will happen is roughly twice as easy as predicting when, and the timing test here is generous: on a twenty-year forecast a 30% window is plus or minus six years, and two-thirds still miss it. Frontier Its standing has to be stated with it. One contractor, one defence sponsor, one scoring rule, one extraction pass. It is the largest such corpus in existence and it is not a population estimate of anything. Speculative The commercial record points the same way from much weaker evidence: Schnaars' Megamistakes is reported, in a newspaper review of it, to have examined dozens of forecasting studies from the 1960s onward and found only about 20% accurate, with the author quoted saying that most technological forecasts “fail not because of mistakes in fine tuning. They fail big.” The book was not obtained for this brief; the 20% figure is second-hand and carried as such.

Established Which leaves the case everyone reaches for, and it is not what it is used for. Gordon Moore's 1965 article in Electronics says complexity at minimum component cost “has increased at a rate of roughly a factor of two per year” and that “over the longer term, the rate of increase is a bit more uncertain, although there is no reason to believe it will not remain nearly constant for at least ten years.” No dataset, two hedges in one sentence, a ten-year horizon, an explicit expiry date. Established Moore revised down to a two-year doubling in 1975 and that proved pessimistic; actual doubling ran near eighteen months through the 1980s and 1990s. In 1994 the Semiconductor Industry Association created a national roadmap, international from 1999, which Chris Mack describes as institutionalising an “industry standard Moore's Law” and supplying “common terminology to suppliers and customers.” Mack, writing from inside the industry, reports that most of it regards the law as “a self-fulfilling prophecy”: “We make Moore's Law happen because we want it to be true.” His verdict is the sentence this brief is built around — “Moore's Law is not a law, it is an act of will.” Established Bloom, Jones, Van Reenen and Webb priced the will: the number of researchers required today to achieve the same doubling of chip density is more than eighteen times the number required in the early 1970s. The output trend held; the input needed to hold it rose more than eighteen-fold. That is not evidence that technological progress is predictable. It is evidence that a sufficiently coordinated industry can hold an exponential for decades at exponentially rising cost.

Frontier Can expert elicitation about arrival dates be made to measure arrival dates? On present evidence it measures something else. Grace and colleagues surveyed 2,778 respondents drawn from authors published in 2022 at the major machine-learning venues. Between the 2022 and 2023 rounds the aggregate 50% date for high-level machine intelligence moved from 2060 to 2047, and the 50% date for full automation of labour from 2164 to 2116 — a 48-year shift in twelve months. Established The framing effect is the real finding. Those two questions describe substantially the same event and their aggregate answers differ by roughly 70 years, a gap that persists across surveys; the authors say it “would seem to be a framing effect,” while allowing that tasks versus occupations may account for part of it. Frontier A method whose answer moves five decades on a rephrasing should be reported with its framing attached and probably as a range spanning both framings, which is not how any such number circulates.

Speculative The most-repeated claim about technology forecasting has had its evidence withdrawn and the withdrawal has not travelled. Armstrong and Sotala assembled a database of 95 AI timeline predictions and reported that they “contradict each other considerably, and are indistinguishable from non-expert predictions and past failed predictions,” clustering at 15 to 25 years out. Established On 11 June 2016 Kaj Sotala published a correction to his own paper. Katja Grace had found a spreadsheet construction and interpretation error; Sotala's statement is that the conclusion that expert predictions “were indistinguishable from those of non-experts” is flawed. He holds that the chapter's broader thesis stands on three other legs — experts disagree widely with each other, many past predictions failed, and the psychological literature suggests expertise here should be very hard to acquire. But the single most-quoted line no longer has its evidence, and the published chapter still carries it. Frontier This is an author correcting against his own interest on a personal blog ten years ago, which is exactly the class of source this site weights most heavily, and it is why the claim is flagged speculative here rather than established.

Frontier Can improvement rates be predicted from the structure of patenting, for domains rather than for products? Benson and Magee regressed patent metrics on measured improvement rates across 28 technological domains, reporting a best single-metric correlation of 0.76 and a multiple regression combining importance, recency and immediacy at R² = 0.64, claimed good for more than ten years ahead. Frontier Carry the correction with the paper. A 2016 correction notice replaces a figure and a table and rewrites a results sentence to read that “there seems to be a slight visual trend in the figure, the Pearson correlation is a moderate 0.38 and the p-value is slightly lower than is generally accepted for statistical significance, at 0.043.” Anyone quoting a correlation from this paper has to say which one. Frontier Triulzi, Alstott and Magee take the same programme further on 30 technologies, finding that the centrality of a domain's patents in the US citation network predicts its improvement rate reliably under Monte Carlo cross-validation once patenting dynamics are normalised out. Speculative The limit is structural and it is the one that matters: the method predicts rates for domains that already have enough patents to be measured, and says nothing about a technology that does not yet exist — which is the class of question people actually pose.

Frontier Where is the horizon break? Two independent datasets agree on the direction and neither supplies a break point. Mullins scores 35% short-term, 28% medium, 27% long; Farmer and Lafond's error grows as the square root of horizon. Speculative Nobody has published the horizon at which a technology forecast stops carrying usable information, and without it the standard practical advice — forecast short — is a direction rather than a rule.

Frontier Are ideas getting harder to find, and if so does every extrapolation overstate the future? Bloom et al. measure research productivity falling sharply across industries, products and firms while research effort rises. If that is general, an extrapolation fitted on a period of cheap discovery systematically overstates the ease of the next increment. Frontier The result is measured and it is contested in the growth literature on measurement grounds — what counts as an idea, what counts as a researcher, and whether the deflators hold — and this brief does not adjudicate that.

Speculative And the synthesis question, which nobody has posed as a hypothesis and which this brief states as its own. If the technologies that get forecast correctly are the ones somebody organised to deliver, then technology forecasting is partly a branch of institutional analysis, and a forecast is partly a report on who is organised. Speculative Nothing in the literature tests this. It is assembled here from Mack's “act of will”, the eighteen-fold input rise, Christensen's disk drives, and the termination of the semiconductor roadmap when the industrial structure that made coordination valuable disappeared. It is flagged as the brief's own construction and should be read as one. Technology Trees of Civilization reaches the same place from the dependency side.

Established The S-curve's central defect is that the parameter you most want is the one you cannot estimate until you no longer need it. Meyer, Yung and Ausubel state it for the logistic: “For incomplete S-shaped processes the estimation of” the saturation ceiling “by OLS depends strongly on small deviations around the midpoint.” Their own recommended workaround is to refit with saturation pinned at, say, 90% of the first fit and compare the two — a sensitivity check, not an estimate. Established Van den Bulte and Lilien measure the same failure in the Bass model: nonlinear least squares estimates are systematically biased, with the coefficient of innovation underestimated by about 20%, the coefficient of imitation underestimated by about 20% and the market ceiling overestimated by about 30%. The cause is ill-conditioning — too little data richness for the model's complexity — intensified by noise, few observations, and right censoring, the gap between the last observation and the true ceiling. Their conclusion is that these problems “make long-term predictive, prescriptive and descriptive applications of Bass-type models problematic.”

Established Put the two together: before the inflection point the ceiling is unidentified, and after it the forecast is nearly trivial. That is the sharpest single statement this brief can make about diffusion forecasting, and it is a mathematical property rather than a criticism of practice. Frontier The standard 25-year review of the diffusion-modelling literature exists and only its bibliographic record was obtainable for this brief; nothing is attributed to it here.

Frontier Technology readiness levels are a maturity scale doing forecasting work they were never built for. Introduced by NASA in the 1970s and now a de facto assessment standard well outside aerospace, a TRL is an ordinal statement about the present. Programme planning routinely converts it into a schedule — a technology at level 4 will reach level 6 by a date — and no measured transition-rate distribution between levels has been published to support that conversion. Olechowski, Eppinger and Joglekar interviewed employees at seven organisations and catalogued 15 challenges in three groups: system complexity, planning and review, and validity of assessment. That third category names the problem exactly. Established The cost side is audited annually: the US Government Accountability Office's 24th annual weapon-systems assessment covers 104 programmes and reports that 18 of 40 programmes entering the middle-tier acquisition pathway between 2018 and 2025 did so with immature technologies, that seven of the eight currently on that pathway still have them, that nine programmes entering with multiple immature technologies “made limited progress in maturing those technologies,” and that average time to deliver capability now exceeds 12 years and is expected to rise. Frontier The audit measures outcomes, not the predictive validity of the scale; the two are routinely conflated. Speculative The originator's own retrospective on technology readiness assessment exists and was not obtained for this brief; it is named here so its absence is not mistaken for its non-existence.

Frontier And the most widely circulated forecasting shape in industry has no supporting data. The leading academic review of the hype cycle reports “incongruences connected with the reports of Gartner” — testing the model against the publishing firm's own data did not reproduce the model's shape — and proposes capturing hype dynamics through existing lifecycle models instead. Handwave Gartner is an interested party: the hype cycle is a commercial product of the firm that publishes it, and the most-cited independent review of it found the vendor's own data incongruent with it. Only the review's bibliographic record and abstract were obtainable here.

Established Nothing in this subject waits on a discovery, and saying so plainly is more useful than manufacturing an edge. No brief on this map produces a result technology forecasting is blocked on. What it waits on is scoring: dated forecasts, resolution criteria fixed in advance, and somebody maintaining the register. None of the three is a research problem and none of them exists. Established The one scored corpus in the field is a fourteen-year-old contractor report; the 2,905 integrated-assessment projections have been scored exactly once, by a group with a competing method to sell; and the tens of thousands of dated technology forecasts published since have never been collected at all.

Frontier There is one genuine missing scientific result and it is narrow enough to name. Nagy et al. showed that Wright's law and Moore's law are nearly indistinguishable in fit because cumulative production is itself close to exponential in time. Nobody has produced an identification strategy that separates learning-by-doing from elapsed time — an instrument, a natural experiment, or a technology whose production history is exogenously non-exponential. Until someone does, the causal claim underneath every deployment-subsidy argument is unidentified, and the honest statement is that the data are consistent with learning-by-doing and equally consistent with its absence. That result is recorded as a typed requirement in section 13.

Established Two further absences are recorded rather than glossed. There is no published transition-rate distribution between technology readiness levels, so the scale cannot be converted into a schedule with a stated error. And there is no population estimate of technology-forecast accuracy: 33% from one contractor and roughly 20% second-hand from a book review are the two largest numbers available, and neither is a sample of anything definable.

Established The cheapest valuable experiment is to score the forecasts that already exist, and the highest-value corpus is sitting in the open. The 2,905 integrated-assessment-model projections of solar cost decline are dated, numerical, published, and resolved by an outturn nobody disputes. They have been scored once, by an interested party. A second scoring by a group with no forecasting method of its own would cost a research assistant a summer and would settle whether the correlated error is a property of one technology, one model family, or the whole practice. Frontier The obvious extension is to repeat it on wind, batteries and electrolysers, where the same models made the same class of projection over the same window.

Established Second: re-run the Mullins extraction under a different sponsor and a pre-registered scoring rule. The 70%/33% split is the field's single most useful pair of numbers and it rests on one contractor's coding decisions about what counts as an event, what counts as realised, and what window counts as correct. Replication would either promote it to a stable finding or reveal it as a scoring artefact, and there is no methodological obstacle to doing it.

Frontier Third: measure the transition rates the readiness scale is used to imply. Programme records at the major defence and space agencies contain dated assessments at successive levels for thousands of technologies. The distribution of time from level 4 to level 6, by domain, is extractable from documents that already exist, and it is the missing input that would turn an ordinal maturity statement into a forecast with an interval. Speculative Whether the resulting distribution would be stable enough to use is genuinely unknown, and the experiment is worth running for the answer either way.

Frontier Fourth, and hardest: find a case where cumulative production is not exponential in time. The Wright-versus-Moore identification problem is a data problem before it is a statistical one. A technology whose output was interrupted — by war, by an export ban, by a supply shock — while its cost series continued would let the two laws be separated. Speculative No such analysis was located for this brief, and the search for suitable cases would itself be a publishable result.

Established Fifth: register forecasts before resolution, in public, with criteria. This is the design the geopolitical forecasting literature already uses and it has never been applied to technology dates at scale. Public Policy Foresight reaches the same recommendation from the institutional side and owns that argument; the version specific to this brief is narrower and cheaper, because a technology forecast usually has a number attached.

Established A rate forecast has hard data requirements and they are why so few technologies have one. The Farmer-Lafond method needs a cost or performance series long enough to estimate drift and autocorrelation — the underlying sample ran 10 to 39 annual observations per technology and 13 of 66 candidate technologies were dropped for lacking a statistically significant improvement rate. Consistent deflation, a stable performance definition and an unbroken series are the whole apparatus, and most technologies have none of them. Established The Santa Fe performance-curve database exists because assembling those series is the expensive part; it is also why the same 50-odd technologies recur across every paper in this literature, which is a sampling fact that should travel with the results.

Established A date forecast needs a threshold definition and this is where most of them fail before they are scored. The Mullins corpus could score only forecasts with an identifiable event and an identifiable date; 2,092 assessable forecasts were extracted from 300 documents and 1,055 survived verification, which is roughly half discarded for reasons that are properties of the writing rather than of the future. Frontier The Grace surveys show what happens when the threshold is defined two ways: a 70-year gap between two descriptions of the same event. An unresolvable forecast is not a bad forecast; it is not a forecast.

Frontier The coordination machinery is engineering too, and it is the part that is usually invisible. The semiconductor roadmap was a document that told hundreds of suppliers which node to build equipment for and when, so that lithography, metrology, materials and design tools arrived together. That is a manufactured schedule, not a prediction, and it required an industry with many mutually dependent firms who could not each build the toolchain alone. Established It ended in 2016. The reason given in the trade press is directly on point: with only a few firms still developing leading-edge fabrication in each of logic and memory, “each leading-edge company has a secret internal roadmap and little motivation to compare directions.” Two successors were announced — a devices-and-systems roadmap in May 2016 and a heterogeneous-integration roadmap — neither with the original's participation. Frontier Technology Trees of Civilization uses the same termination for a different purpose, as the fate of the most successful roadmap ever published; here it is the moment the industry's forecasting apparatus stopped being worth maintaining.

The boundary with Public Policy Foresight is the reason this brief exists and it is stated rather than implied. That brief owns futures methodology as a practice: government foresight units, Delphi and its comparative record, scenario planning, horizon scanning, national risk registers, the geopolitical forecasting tournaments and the re-analyses of them, and the central finding that the field decided outcome measurement was the wrong criterion and never agreed a replacement. None of that is re-derived here. This brief owns the parts that literature does not carry: S-curves and their identification failure, Wright's law and experience curves, Moore's law as a coordination target rather than a forecast, bibliometric and patent-based prediction of improvement rates, and the tracked record of dated technology forecasts. Three neighbouring slots — futures methodologies, scenario planning and emerging-technology assessment — were cross-listed to that brief and to Scientific Advisory Institutions rather than authored, because they are one literature with four vocabularies. This one was kept because it is not.

Within this map, also: Technology Trees of Civilization, which owns dependency structure and roadmapping and shares exactly two sources with this brief for opposite purposes; Civilization Timelines, which owns long-run growth models as stage theories where this brief owns extrapolation methods; Megaproject Governance, which owns cost and schedule overrun on projects rather than on technologies; Innovation Ecosystems and Scientific Funding Models, where the falling research productivity measured here is a funding-design problem; Artificial General Intelligence, which owns AI capability forecasting as a subject rather than as an example; Energy Storage Revolutions and Space-Based Solar Power, whose cost trajectories are among the series these methods are fitted to; and Civilizational Planning, which inherits the question of what a plan built on a forecast is worth.

Outside it: the economics of technical change, which supplies the experience-curve tradition; industrial organisation, which supplies the semiconductor case; scientometrics, which supplies the patent-network methods; marketing science, which is where the diffusion-model identification problem was actually diagnosed; and applied statistics, which supplies the hindcasting design.

Established The distinctive institutional fact about this subject is that nobody owns the scorecard. Forecasts are produced by consultancies, trade bodies, government offices, integrated assessment modellers and individual researchers; scoring them is nobody's mandate, nobody's budget line, and in several cases directly against the producer's interest. The one large scoring exercise in the record was commissioned by a defence sponsor and has not been repeated in fourteen years. Frontier That is a weaker version of the finding Public Policy Foresight reaches about foresight institutions, arrived at in a different literature, and the two should be read together rather than as independent confirmations.

Established Interested parties are unusually easy to name here and it matters which way the interest runs. Moore was Fairchild's research director when he wrote the 1965 article and later co-founded the firm with most to gain from the trend continuing. Mack is a lithography consultant, which is why his admission that the law is an act of will carries weight — it costs him something. Way et al. criticise the modelling community while promoting their own method. The authors of the logistic-fitting primer also distribute the software it documents. Gartner sells the hype cycle. Established Two sources in this brief run against their authors' interest and are weighted more heavily for it: Wright's own admission that his curve came from two or three points, and Sotala's public withdrawal of the finding his own chapter is most cited for.

Frontier The institutional requirement that would change the subject is a register. Dated forecasts, stated resolution criteria, a fixed scoring rule, and a body that keeps it running past the tenure of whoever started it. Nothing about that is technically hard; it has simply never been anyone's job. Established Both of the institutional constraints recorded in section 13 are of that kind — things a funder or a standards body could choose to supply and has not.

Established A widely believed technology forecast is not an observation of the world; it is an input to it, and the documented case runs in the losing direction. Christensen's disk-drive history found that individual component improvements did follow S-curves, but that “the flattening of S-curves is a firm-specific, rather than uniform industry phenomenon,” and — the load-bearing sentence — that “lack of progress in conventional technologies may be the result, rather than the stimulus, of a forecast that the conventional technology is maturing.” Established A firm forecasts maturity, stops investing, and observes maturity. He also found that firms pursuing aggressive S-curve switching gained no strategic advantage over firms extending established component technologies, which removes the obvious remedy. Frontier Set that beside Moore's law and the pair is the cleanest argument in the brief: the same mechanism runs both ways, and a forecast published with authority is a claim on other people's investment decisions.

Established The 2,905-projection error had consequences and they were not symmetric. Projections that solar costs would fall 2.6% a year, produced by the modelling community that advises governments on energy, fed cost-benefit analyses of decarbonisation over the decade in which costs actually fell 15% a year. Frontier Attributing specific policy outcomes to that error is beyond what the evidence supports and this brief does not do it. What is established is that the error was unidirectional, professional, and repeated across nearly three thousand published numbers.

Established Publishing a point forecast without its interval is the routine failure and it is an ethical one, not a technical one. The one validated method in the field produces a distribution; the number that reaches a decision-maker is a line on a chart. Frontier The experience curve is the sharpest case: an 80% learning rate quoted as a manufacturing constant is a firm-level exponent from 1936 being used to underwrite a subsidy programme, and the person quoting it is usually unaware that its author called it a presumption from two or three points.

Established The terminal position of this brief is that the field's canonical success is not evidence of what it is used to prove. Moore's law held for fifty years. The industry that held it built an institution in 1994 to keep it true, describes it as a self-fulfilling prophecy, and calls it an act of will. The input cost of holding it rose more than eighteen-fold for the same output. And when consolidation removed the reason for firms to coordinate, the institution shut down. The most-cited evidence that technological progress is predictable is instead evidence that a sufficiently coordinated industry can make a prediction come true — and, via Christensen, that a sufficiently discouraged firm can make one come true in the other direction. That is a finding about organisation, not about foresight.

Frontier The corollary is uncomfortable for any map of frontier technology, including this one. If the technologies that get forecast correctly are the ones somebody organised to deliver, then a technology forecast is partly a report on who is organised, and the interesting variable is institutional rather than physical. Technology Trees of Civilization arrives at the same conclusion from the dependency side, measuring that the constraints binding frontier subjects are overwhelmingly institutional rather than technological. Two briefs, two methods, one conclusion, and neither of them set out to reach it.

Established What survives as usable is narrower than the field's name and more valuable than its reputation. For a technology already on a measured cost trend, the expected magnitude of a forecast error at a given horizon can be stated, and that statement has been validated on thousands of hindcasts. That is a real instrument. It does not answer when a capability will exist, it does not work on technologies that do not yet have a series, and it produces intervals rather than dates. Frontier A civilisation planning around technology should ask for the interval and should treat any date without one as a statement of intent.

Speculative And the long-run question is whether the conditions that made the one success possible recur. Many mutually dependent firms, a common technical schedule, no single actor able to build the toolchain alone, and enough rivalry to make falling behind expensive. Those conditions produced a fifty-year exponential and then dissolved. Whether any comparable structure exists in energy, in biotechnology or in computing's successor is not a forecasting question and nobody has posed it as an empirical one.

These horizons deliberately give no arrival dates. The brief's own evidence says that dated technology forecasts resolve inside a generous window about a third of the time, so what follows tracks methods, corpora and institutions instead:

• 10 yr: Frontier The methodological picture is unlikely to change, because it is not blocked on anything. Expect more hindcast validation of correlated-random-walk models on more technology series, and expect the patent-network improvement-rate methods to be extended to more domains. Speculative The scoring gap is the one thing that could close quickly and cheaply — the 2,905 integrated-assessment projections are scorable by anyone with a spreadsheet — and it has stayed open for four years since the outturn became unambiguous. Frontier The two semiconductor roadmap successors either acquire the participation the original had or they do not, and that is observable within this window without forecasting anything.
• 25 yr: Speculative If a public register of dated forecasts with resolution criteria is ever built, the first defensible population estimate of technology-forecast accuracy becomes possible about a decade after it starts, and not before. Speculative If it is not built, the field's accuracy figures in 2050 will still be a 2012 contractor report and a 1989 book review. Handwave Which of those happens is a funding decision, and predicting a funding decision is not extrapolating a measurement.
• 50 yr: Speculative The deep open question is whether falling research productivity is a general law or an artefact of measurement. If it is general, every extrapolation fitted on the twentieth century overstates the twenty-first, and the experience-curve literature's headline exponents drift. Speculative If it is measurement, they do not. Frontier The two readings are both fitted to real data and this brief declares the tie rather than choosing.
• 100 / 250+ yr: Handwave Beyond any evidence this brief has. The longest validated forecast horizon in the record is about twenty years, the error at that horizon is already large, and it grows as the square root of the horizon — so a century-scale technology forecast is an assertion with a number attached. Handwave The only durable thing to say is structural: the technologies that arrived on schedule were the ones an organised industry had committed to deliver, and nothing about that observation extrapolates.

• Depends on Nothing on this map. No brief here produces a result technology forecasting is waiting on: the methods are published, the mathematics is settled, and what is missing is measurement infrastructure and one identification. No typed depends-on edge is claimed.
• Requires (not on this map) One scientific result and two institutional facts, none of them produced by any brief on this map. The scientific one is specific and a researcher in the field would recognise it immediately: Wright's law and Moore's law fit 62 technologies almost equally well because cumulative production is itself close to exponential in time, so the comparison cannot say whether cost falls because output accumulates or because time passes. What is missing is an identification strategy — an instrument, a natural experiment, or a technology whose production history was exogenously interrupted while its cost series continued — that separates learning-by-doing from elapsed time. Until it exists, the causal premise under every deployment-subsidy argument is unidentified, and this is a discovery nobody is currently attempting rather than an engineering task. The institutional ones are cheaper and even more conspicuously absent. There is no maintained public register of dated technology forecasts with resolution criteria fixed in advance, so the field's accuracy figures rest on a single 2012 contractor report scoring 1,055 forecasts under one sponsor's coding rules. And the 2,905 integrated-assessment-model projections of solar cost decline over 2010 to 2020 — mean 2.6% a year, none above 6%, against a 15% outturn — have been scored exactly once, by the research group that publishes the competing method. A second scoring by anyone with nothing to sell would cost a summer of one person's time. Neither institutional requirement is a research problem; both are things a funder or a standards body could choose to supply and has not.
• Enables In principle every cost trajectory this map quotes — storage, solar, electrolysis, launch — is the output of one of these methods, and the intervals around those trajectories come from this literature or from nowhere. No typed enabling edge is claimed, because a method is not a result another brief waits on, and recording an edge here to improve a coverage statistic is precisely the failure Technology Trees of Civilization measures.
• Adjacent The economics of technical change; scientometrics and patent analysis; marketing science, where the diffusion-model identification problem was diagnosed; applied statistics for the hindcasting design; industrial organisation for the semiconductor case; and within this map Public Policy Foresight, Technology Trees of Civilization and Megaproject Governance.

Established “The 80% learning curve is a constant of manufacturing.” It is one firm's exponent, published in 1936, and its author says in the same paper that it was “worked up empirically from the two or three points which previous production experience of the same model in differing quantities made possible.” Frontier Wright's relationship has since been confirmed very broadly — that is not in dispute — but learning rates vary widely by technology, and quoting 80% as a default is quoting a number that was never a measurement of anything general.

Speculative “Experts are no better than non-experts at predicting AI.” The study everyone cites for this had the supporting analysis withdrawn by its own author in June 2016 after an error was found in the underlying spreadsheet. Established The published chapter still carries the sentence; the correction lives on a personal blog. The chapter's other legs — that expert predictions disagree widely with each other and that many past predictions failed — are unaffected and are separately supported. Anyone citing the indistinguishability finding without the correction is repeating a known error.

Established “Moore's law proves exponential technological progress is normal.” One industry, one metric, one deliberately created coordinating institution, a self-description as a self-fulfilling prophecy from inside that industry, an eighteen-fold rise in the researchers needed for the same doubling, and a roadmap that its own participants terminated in 2016. Frontier Every element of that sentence is measured, and together they describe an achievement of industrial organisation rather than a regularity of technology. The inference from “Moore's law held” to “progress is exponential” discards all of it.

Handwave “The hype cycle describes how technologies are received.” It is a commercial product of the firm that publishes it, and the most-cited academic review of it reports incongruence with that firm's own published data. Speculative No independent dataset reproducing the curve was located for this brief. Using it as a diagnostic is using a shape for which the vendor's own numbers are the disconfirming evidence.

Frontier “Wright's law shows that deployment subsidies cut costs.” It shows that cost falls as cumulative production rises. Because production grows roughly exponentially in time, a model with no mechanism at all fits almost as well, and the comparison therefore cannot identify the mechanism. Speculative The data are consistent with learning-by-doing and equally consistent with its absence, and the policy argument needs the causal reading that the evidence does not supply.

Frontier “S-curve analysis tells you when a technology is maturing.” Before the inflection point the ceiling is not identifiable from the data; Bass-model parameters are biased by about 20%, 20% and 30% in known directions, and right censoring makes it worse. Established And the diagnosis is not inert: Christensen documented firms whose forecast of maturity produced the maturity, with no strategic advantage to the firms that switched. Read as a planning tool rather than as a description, the S-curve is a mechanism for manufacturing the outcome it predicts.

Frontier “A technology readiness level tells you how long the technology needs.” It is an ordinal statement about the present with no published transition-rate distribution attached. Established The audit record shows what the conversion costs when it is made anyway: 18 of 40 programmes entering a fast-track acquisition pathway with immature technologies, seven of eight still immature, and average delivery time above twelve years and rising. That measures outcomes, not the scale's predictive validity, and the two are not the same claim.

Established “Technology forecasts are about a third accurate.” Two numbers circulate for this — 33% and roughly 20% — and neither is a population estimate. The first is one contractor's scoring of 1,055 forecasts under one sponsor for one set of coding rules; the second is second-hand from a newspaper review of a book this brief did not obtain. Speculative There is no defensible general accuracy rate for technology forecasting, and the honest statement is that the measurement has not been taken.

Frontier And the framing itself: “can technology be forecast” is not one question. Ask about the improvement rate of a technology already on a measured trend and the answer is yes, with a validated error distribution. Ask when something will exist and the answer is a third of the time, generously scored. Ask how far it will spread before it saturates and the answer is that the parameter you want is unidentified until you no longer need it. Established The field's credibility floats between the three because one phrase covers all of them.