Overview and size of the alternative data market.

What alternative data is

Lesson 5.2.2 defined it by exclusion: everything that is not prices, volumes or financial statements. It is a useful definition for a model, but it does not say where the data comes from or why anyone sells it. A more practical classification is by origin, because the origin determines the coverage, the delay and the biases (lesson 6.7.2) and also the legal risks (lessons 6.7.3 and 6.7.4):

flowchart LR
  A[Alternative data] --> P[People]
  A --> E[Companies]
  A --> S[Sensors]
  P --> P1[Cards and receipts]
  P --> P2[Websites and reviews]
  E --> E1[Jobs and patents]
  E --> E2[Published prices]
  S --> S1[Satellite and AIS]
  S --> S2[Weather and location]

The level 6 data includes a synthetic example of almost every row, all of them about the 120 fictitious companies in the level 5 panel: cards (25 companies), web and app traffic (30), employment (60), reviews (10), ships (10 exporters), weather (12 utilities), parking lot images (8 chains) and patents (40). Each source measures part of the demand shock that moves sales relative to consensus, with its own noise and delay; lesson 6.7.2 goes through them one by one.

The value chain

Between the event (someone pays by card in a store) and a fund's signal there are several links, and each one charges:

flowchart TB
  F[Source] --> V[Vendor]
  V --> C[Catalog]
  C --> B[Buyer]
  B --> S[Signal]

• Source: whoever generates the trace (a bank, an app, a satellite, a website). It often does not know that its trace has value for an investor.
• Vendor: buys or collects the trace, cleans it, aggregates it, maps it to companies and sells it with a license (lesson 1.3.7). It is the one that bears the legal risk of the collection.
• Catalog or aggregator: companies such as Neudata or Eagle Alpha maintain catalogs of datasets with profiles, indicative prices and evaluations; platforms such as Nasdaq Data Link distribute third-party datasets. Lesson 6.7.4 covers their role in due diligence.
• Buyer: quantitative and discretionary funds, above all; also banks, insurers and companies that keep an eye on their competitors.
• Signal: the transformation of the data into a prediction that can be traded (lesson 6.7.5).

This chain explains why “market size” figures differ so much: you can measure what buyers of one type (investment managers) spend, what vendors earn from all their clients, or a “market” that includes data collected inside each company. Each choice gives a different figure.

What Neudata measures and how

For years Neudata has published an estimate of investment managers' spending on alternative data. Its figures can be read on its website11Note 1 · Neudata, “How big is the alternative data market for investment managers?” (https://www.neudata.co/education/how-big-is-the-alternative-data-market-for-investment-managers, accessed on 2026-10-23; the page is undated and refers to the 2025 report): “we included 2,338 datasets in our analysis”; “((max−min)×0.3)+min”; “if a dataset was rated as having 6-10 investment clients, we assumed eight distinct buyers”; growth of 15%, 13%, 26% and 33% from 2021 to 2024; “the average dataset (as of 2024) earns $1.1m/yr”; sensitivities of 11.4 billion (Lite profiles) and of 2.7 to 12.7 billion (catalog = 90%); “the margin for error is large” and “a proxy for alternative data spending rather than a precise measure”.22Note 2 · Neudata, press release “Alternative data spending by investment management firms could top $15.4bn in 2025”, London, 2025-02-26 (https://www.neudata.co/press-release/alternative-data-spending-by-investment-management-firms-could-top-15-4bn-in-2025-according-to-neudata-s-latest-report, accessed on 2026-10-23): “a minimum of $2.5 billion […] in 2024”, average growth of 21% a year since 2020, 33% in 2024, “nearly 8,000 unique datasets” in its catalog, and from 8.6 to 39.9 billion in 2030.33Note 3 · Neudata, “Alternative data market trends in 2026” (https://www.neudata.co/blog/state-of-the-alternative-data-market-2026, 2026-02-24, accessed on 2026-10-23): “approximately $2.8bn in 2025, growing 17% year-on-year”; “2,805 datasets in 2025, up from 2,215 in 2024”; “approximately $23.1bn by 2030 if current growth trends persist”. The page does not give the base of that projection.:

The method is described in the 2025 article¹: for each dataset in the catalog, the annual price given by the vendor times the number of investment manager clients it has. With three explicit assumptions:

• If the vendor gives a price range, the midpoint is not used but rather the value at 30% of the range, ((max − min) × 0.3) + min, because actual transactions are concentrated below the midpoint.
• If the number of clients comes in bands, the midpoint is taken (from 6 to 10 clients, 8 buyers).
• Only datasets with a full profile count. If the 5,600 short profiles (Lite) were added with the average spending of 1.6 million, the 2024 figure would go from 2.5 to 11.4 billion; and if the catalog were 90% of total spending, the 2024 range would go from 2.7 to 12.7 billion.

Neudata calls it “a proxy for alternative data spending rather than a precise measure”, and warns that vendors may give it a list price that is later negotiated down and that “the margin for error is large”. With that description, the figures can be reconstructed:

import numpy as np
import pandas as pd

crec = pd.Series({2021: 0.15, 2022: 0.13, 2023: 0.26,
                  2024: 0.33, 2025: 0.17}, name="crecimiento")
acum = (1 + crec.loc[2021:2024]).prod()
media = acum ** (1 / 4) - 1
print(f"2021-2024: x{acum:.3f}, geometric mean {media:.1%}")
base_2020 = 2.5 / acum
print(f"implied spending in 2020: {base_2020:.2f} billion")
nivel = base_2020 * (1 + crec).cumprod()
print(nivel.round(2).to_dict())

2021-2024: x2.178, geometric mean 21.5%
implied spending in 2020: 1.15 billion
{2021: 1.32, 2022: 1.49, 2023: 1.88, 2024: 2.5, 2025: 2.92}

The four growth rates given in the article match the 21% average announced in the press release. The 2020 spending they imply is not published by Neudata: it is a derived figure, and it has to be presented as such. The last row mixes two reports: the 17% for 2025 comes from the 2026 report, with a different sample of datasets, and applied to the 2.5 billion it gives 2.92 billion, not the 2.8 billion that report publishes (next section).

The 2025 and 2030 ranges in the 2025 report are reconstructed the same way: the 2024 sensitivity range (from 2.7 to 12.7 billion) grown at 21% a year. The cell checks it:

r24 = np.array([2.7, 12.7])     # 2024, Neudata sensitivity
g = 0.21
r25 = np.round(r24 * (1 + g), 1)
r30 = np.round(r25 * (1 + g) ** 5, 1)
print("2025 reconstructed:", r25, "published: [ 3.3 15.4]")
print("2030 reconstructed:", r30, "published: [ 8.6 39.9]")
print(f"from 2025 to 2030, x{1.21 ** 5:.2f}")

2025 reconstructed: [ 3.3 15.4] published: [ 3.3 15.4]
2030 reconstructed: [ 8.6 39.9] published: [ 8.6 39.9]
from 2025 to 2030, x2.59

The reconstruction matches what was published (2025 is the 2024 range times 1.21, rounded; 2030, the 2025 range times $1.21^5 = 2.59$). This says two things: the ranges are not independent forecasts, but a mechanical extrapolation of past growth; and the width of the range (from 1 to 4.7) does not come from uncertainty about the future, but from uncertainty about the present (what share of spending the catalog sees).

The method, applied to a made-up catalog

To see how much the assumptions move the result, the following cell applies Neudata's rules to a made-up catalog of 2,338 datasets (the number in the report), with price ranges and client bands generated with a fixed seed. The absolute figures mean nothing; the ratios between rules do:

rng = np.random.default_rng(671)
n = 2338
minimo = rng.choice([25, 50, 100, 150, 250], n) * 1e3
maximo = minimo * rng.choice([2, 3, 4], n)
tramos = np.array([(1, 5), (6, 10), (11, 25), (26, 50)])
t = tramos[rng.choice(4, n, p=[0.45, 0.25, 0.2, 0.1])]

def total(precio, clientes):
    return (precio * clientes).sum() / 1e9

reglas = {
    "midpoint": total((minimo + maximo) / 2, t.mean(1)),
    "Neudata (30%)": total(minimo + 0.3 * (maximo - minimo),
                           t.mean(1)),
    "minimums": total(minimo, t[:, 0]),
}
ref = reglas["Neudata (30%)"]
for k, v in reglas.items():
    print(f"{k:15s} {v:5.2f} billion  ({v / ref:.2f}x)")

midpoint         5.87 billion  (1.25x)
Neudata (30%)    4.71 billion  (1.00x)
minimums         1.87 billion  (0.40x)

Switching from the midpoint to the 30% rule lowers the total by a fifth in this catalog; taking the minimum price and the minimum number of clients leaves it at 40% of Neudata's figure. None of the three is “the truth”: the truth is private contracts that nobody publishes. That is why a market figure without its method is not a data point, it is a claim.

Revisions between reports

The 2026 report gives 2.8 billion for 2025 with growth of 17%³. The 2024 figure it implies is not the one in the previous report, and the 2030 figure does not follow from the rate it announces:

g25, gasto25, meta30 = 0.17, 2.8, 23.1
print(f"implied 2024: {gasto25 / (1 + g25):.2f} (published "
      "in 2025: at least 2.5)")
tasa = (meta30 / gasto25) ** (1 / 5) - 1
print(f"annual rate from 2.8 to 23.1 in five years: {tasa:.1%}")
for g in (0.17, 0.21):
    print(f"2030 at {g:.0%} a year from 2.8: "
          f"{gasto25 * (1 + g) ** 5:.1f}")
print(f"2025 base that gives 23.1 at 21%: "
      f"{meta30 / 1.21 ** 5:.1f}")

implied 2024: 2.39 (published in 2025: at least 2.5)
annual rate from 2.8 to 23.1 in five years: 52.5%
2030 at 17% a year from 2.8: 6.1
2030 at 21% a year from 2.8: 7.3
2025 base that gives 23.1 at 21%: 8.9

With 2.8 billion and 17% a year, 2030 would come to 6.1 billion; with 21%, to 7.3 billion. The 23.1 billion requires a rate of 52.5% a year from 2.8, or starting in 2025 from a base of 8.9 billion at 21%. The public pages of the report do not say what the base of that projection is (the 2025 report worked with a range of bases); the full report is downloaded after filling in a form. It is a question to ask the vendor before citing the figure, not a figure to repeat.

The sample also changes: the 2025 article analyzes 2,338 datasets in 2024; the 2026 blog post speaks of 2,215 in 2024 and 2,805 in 2025. Comparing the spending in two reports means comparing two different catalogs.

The consultancies: widely differing estimates

Market research firms publish their own “alternative data markets”. The public part of each report (the press release or the product page) gives one or two figures, a base year, a compound annual rate and a period. Those of five firms, read on their websites on 2026-10-2344Note 4 · Pages accessed on 2026-10-23: Grand View Research (https://www.grandviewresearch.com/press-release/global-alternative-data-market, “October 2024”); IMARC (https://www.imarcgroup.com/alternative-data-market, “Last Updated: October 05, 2026”); Technavio (https://www.technavio.com/report/alternative-data-market-industry-analysis, “Published: Jan 2025”); Precedence Research (https://www.precedenceresearch.com/alternative-data-market, “Last Updated: 17 Sep 2026”); Arizton (https://www.arizton.com/market-reports/alternative-data-market, published in November 2024 and updated in February 2026). Only the figures from the public part of each page are used.:

To compare them you have to bring them to the same years. With the rate and the period each firm gives, the missing values are filled in (the derived values, with their origin):

def compuesto(v, g, n):
    return v * (1 + g) ** n

base_tn = 60.32 / (1.525 ** 5 - 1)   # Technavio: 2024 base
est = pd.DataFrame([
    ("Neudata 2026", 2.8, 23.1),
    ("Neudata 2025 low", 3.3, 8.6),
    ("Neudata 2025 high", 15.4, 39.9),
    ("Grand View", 135.72 / 1.634 ** 5, 135.72),
    ("IMARC", 12.0, compuesto(16.2, 0.34, 4)),
    ("Technavio", base_tn * 1.525, np.nan),
    ("Precedence", 14.16, compuesto(21.61, 0.5068, 4)),
    ("Arizton", compuesto(6.27, 0.5261, 2),
     compuesto(6.27, 0.5261, 7)),
], columns=["firma", "y2025", "y2030"]).set_index("firma")
est.round(1)

                   y2025  y2030
firma
Neudata 2026         2.8   23.1
Neudata 2025 low     3.3    8.6
Neudata 2025 high   15.4   39.9
Grand View          11.7  135.7
IMARC               12.0   52.2
Technavio           12.7    NaN
Precedence          14.2  111.4
Arizton             14.6  120.9

Grand View does not publish the 2025 value on the page (it comes from discounting its 2030 figure at its rate); the 2030 value for IMARC, Precedence and Arizton comes from applying their rate to their last published value; and Technavio gives an increase, not a level, so its 2024 base is solved from the rate (with its period from 2024 to 2029 there are five years of growth; the page itself also mentions 2025-2029, which would give another base). Before using any of these figures, it is worth checking whether they are consistent with themselves:

print("IMARC 2034 reconstructed:",
      round(compuesto(16.2, 0.34, 8), 1), "(published 168.2)")
print("Precedence 2035 reconstructed:",
      round(compuesto(21.61, 0.5068, 9), 1), "(published 854.34)")
print(f"Precedence, from 2025 to 2026: {21.61 / 14.16 - 1:.1%}")
consult = est.drop(index=[i for i in est.index
                          if i.startswith("Neudata")])
for col in ["y2025", "y2030"]:
    v = consult[col].dropna()
    nd = est.loc["Neudata 2026", col]
    print(col, f"consultancies from {v.min():.1f} to {v.max():.1f}",
          f"(x{v.max() / v.min():.1f});",
          f"versus Neudata 2026: x{v.min() / nd:.1f}"
          f" to x{v.max() / nd:.1f}")

IMARC 2034 reconstructed: 168.4 (published 168.2)
Precedence 2035 reconstructed: 865.3 (published 854.34)
Precedence, from 2025 to 2026: 52.6%
y2025 consultancies from 11.7 to 14.6 (x1.3); versus Neudata 2026: x4.2 to x5.2
y2030 consultancies from 52.2 to 135.7 (x2.6); versus Neudata 2026: x2.3 to x5.9

The figure at the top plots the table: each row goes from 2025 to 2030 on a logarithmic scale, and the only highlighted point is the spending Neudata estimates for 2025. The five consultancies agree fairly well on 2025 (from 11.7 to 14.6 billion) and diverge in 2030 (from 52.2 to 135.7 billion), because their rates range from 34% to 63% a year. Three reasons explain the gap with Neudata:

1. What they measure. Neudata, the spending of investment managers on datasets in its catalog. The consultancies, a market with other buyers: their public segmentation includes banking and insurance, retail and other sectors, in addition to funds. Grand View attributes 68.0% of 2024 revenue to hedge funds; even so, its funds market would be several times the total spending of investment managers according to Neudata.
2. How they measure it. Neudata publishes its method and its assumptions. The consultancies' public pages do not: the method, if it exists in writing, is in the paid report.
3. What rate they assume. 50% a year multiplies the market by 7.6 in five years and by 58 in ten. The growth measured by Neudata between 2021 and 2025 was 13% to 33% a year. A sustained rate of 50% for a decade is an extraordinary hypothesis, and none of the pages justifies it.

The figure that matters to the buyer

For a fund, the size of the market is context. What decides is whether a specific dataset pays for itself. Neudata estimates that the average dataset in its catalog earned 1.1 million dollars a year from its investment manager clients in 2024¹; in addition to the price there is an internal cost of ingestion, cleaning, mapping to companies and research that the lesson cannot measure with public sources. The cell assumes 0.9 million (an assumption of the course, not a data point) and computes how much the dataset has to earn, in basis points on the portfolio, to cover both costs:

precio, interno = 1.1e6, 0.9e6   # internal: an assumption of the course
cartera = pd.Series([0.1, 0.5, 1, 5, 20, 50], name="cartera_mm")
aum = cartera.to_numpy() * 1e9
umbral = pd.DataFrame({
    "solo_precio_pb": precio / aum * 1e4,
    "con_interno_pb": (precio + interno) / aum * 1e4,
}, index=cartera)
umbral.round(1)

            solo_precio_pb  con_interno_pb
cartera_mm
0.1                  110.0           200.0
0.5                   22.0            40.0
1.0                   11.0            20.0
5.0                    2.2             4.0
20.0                   0.6             1.0
50.0                   0.2             0.4

A 100 million fund needs 200 bp a year just to cover the cost, something almost no dataset delivers to the whole portfolio. A 20 billion one needs 1 bp. This is the economy of scale that explains who buys: expensive data is concentrated in large portfolios and in strategies that use the same data in many positions. For an individual or a small fund, the realistic path is their own cheap data (public sources, responsible scraping, lesson 6.7.3) or datasets priced by usage.

The other side of the calculation is decay: the more buyers a dataset has, the sooner its information is reflected in the price. Neudata reports that in 2025 the typical dataset had about 20 investment manager clients, versus 25 in 202455Note 5 · Neudata, press release “AI adoption is rising, but alt-data vendor spend hasn't” (London, 2026-02-23, https://www.neudata.co/press-release/ai-adoption-is-rising-but-alt-data-vendor-spend-hasn-t-neudata-report-finds, accessed on 2026-10-23): “around 20 investment clients on average, down from 25 in 2024”., and interprets it as fragmentation of the market, not saturation. In the course data the decay is planted: the card signal loses almost all its value once its use is “published” in 2021 (alt6_verdad_trimestral.csv, column phi; lesson 6.7.5).

Common mistakes

• Citing a market figure without a date or definition. The 2025 figures range from 2.8 to 14.6 billion depending on who measures what. Without a definition, it cannot be compared with anything.
• Taking an extrapolation for a forecast. Neudata's ranges for 2030 are the 2024 range grown at 21% a year; the consultancies' figures, a rate applied to a base. None of them is a model of the market.
• Mixing reports. The 2024 figure in the 2025 report (at least 2.5 billion) and the one implied by the 2026 report (2.39 billion) come from different catalogs. A growth rate computed between them means nothing.
• Confusing spending with value. That a lot is spent on a type of data does not mean it works: Neudata counts what is paid, not what is earned. Lessons 6.7.4 and 6.7.5 cover how a dataset is evaluated.
• Ignoring the internal cost. The license price is only one part: ingestion, mapping to companies, quality control and research cost as much or more, and they do not scale with the size of the portfolio.

Exercises

1. Neudata gives a minimum spending of 2.5 billion in 2024, 11.4 billion if the short profiles are added and a range of 2.7 to 12.7 billion if the catalog is 90% of spending. Reconstruct the range and explain the difference at the low end.

for v in (2.5, 11.4):
    print(f"{v} / 0.9 = {v / 0.9:.2f}")
v = np.arange(2400, 2600) / 1000
ok = v[(np.round(v, 1) == 2.5) & (np.round(v / 0.9, 1) == 2.7)]
print(f"rounds to 2.5 and gives 2.7: from {ok.min()} to {ok.max()}")

2.5 / 0.9 = 2.78
11.4 / 0.9 = 12.67
rounds to 2.5 and gives 2.7: from 2.451 to 2.474

2. A vendor offers you a dataset for 400,000 dollars a year and estimates the cost of integrating it at 300,000. Your fund manages 2 billion, but the signal only affects a third of the portfolio (the stocks the dataset covers). How many basis points does it have to deliver on that third to pay for itself?

coste = 0.4e6 + 0.3e6
cubierta = 2e9 / 3
print(f"on the whole portfolio: {coste / 2e9 * 1e4:.1f} bp")
print(f"on the covered part: {coste / cubierta * 1e4:.1f} bp")

on the whole portfolio: 3.5 bp
on the covered part: 10.5 bp

3. Pick an alternative data market report that is not in the table. Note what it measures, the base year, the rate, the period and whether it publishes its method. How many times is the market multiplied over its horizon at that rate? Is it compatible with the growth measured by Neudata?