---
title: "Alternative Data Strategies"
book: "Strategies I: Equities and Futures"
subject: quant
language: en
chapter: 16
exercises: 8
source: https://one-course.com/books/quant/8/en/chapter/16-alternative-data-strategies
---

# Chapter 16 — Alternative Data Strategies

When satellite images of retailers’ car parks first reached investors, those who bought them could trade ahead of the retailers’ quarterly reports, especially the bad ones; Katona, Painter, Patatoukas and Zeng found more informed short selling and lower liquidity around the reports of covered retailers, and concluded that unequal access to such data increased information asymmetry without immediately improving price discovery. An alternative dataset is worth most to its first buyers and less to each one after. On the synthetic market a panel that reads a tenth of the companies’ earnings surprises ten days early earns a Sharpe ratio of 4.4 when no one else has it, 1.9 when half the market does, and loses when nine tenths do; a firm that gets the same data one day before everyone else keeps 4.3 whatever the diffusion. This chapter takes a dataset to a position and measures how its value decays as it spreads. The build is `firm.altstrat`.

## 16.1 From dataset to position

**Definition 16.1 (Nowcast, alternative-data strategy).**

A *nowcast* is an estimate of a quantity that will be reported later (sales, earnings, economic activity) made from data observed before the report. An *alternative-data strategy* trades a security on nowcasts built from data outside the usual market and accounting sources (transactions, images, web traffic, text), ahead of the report that will reveal the quantity.

The path from a dataset to a position has four steps. Map the data to the companies it measures (Book 7, chapter 12 discussed coverage and the point-in-time trap of backfilled histories). Turn the readings into a [nowcast](#def-s1-alternative-data-strategies-nowcast) of the reported quantity: here, the earnings surprise. Turn the [nowcast](#def-s1-alternative-data-strategies-nowcast) into an expected return: the announcement’s price reaction per unit of surprise. And hold the position over the window in which the market learns what the data already say. The chapter’s panel covers a tenth of the synthetic companies (about 420 announcements a year), reads each surprise with noise twice its spread (a correlation of 0.46), and is delivered ten trading days before the announcement. The [nowcast](#def-s1-alternative-data-strategies-nowcast) is the reading times a slope of 0.21 fitted on the first two years; the book holds the [nowcast](#def-s1-alternative-data-strategies-nowcast) from the delivery’s close to the announcement’s close, scaled to a gross exposure of one.

## 16.2 What has been documented to work

The published record is a set of datasets, each tied to a quantity that prices learn slowly. Satellite coverage of retailers let sophisticated investors target bad reports (Katona and co-authors). Crowdsourced employer reviews forecast earnings surprises: Green, Huang, Wen and Zhou found that firms whose ratings improved significantly outperformed firms whose ratings fell, with the effect in current employees’ views of career opportunities and senior management. Economic links are an older example: Cohen and Frazzini found that prices do not promptly incorporate news about a firm’s principal customers, and that a strategy trading suppliers on their customers’ returns earned monthly alphas of over 150 basis points; the synthetic market’s supplier links (Book 7, chapter 10) plant that effect. In each case the edge is the gap between when the data know and when the price does.

## 16.3 Decay as the data spread

**Definition 16.2 (Data decay).**

*Data decay* is the loss of a dataset’s trading value as more investors acquire it and trade on it earlier, moving prices before the report and leaving less of the reaction for each user; unlike the post-publication decay of a published anomaly, it is driven by the number of buyers of the data.

Zhu found the other side of it: the introduction of consumer-transaction and satellite data raised stock price informativeness by lowering the cost of acquiring information. More informative prices are good for markets and bad for the dataset’s value. The chapter plants the mechanism directly ([Listing 16.1](#lst-s1-alternative-data-strategies-diffuse)): a share $\phi$ of the market buys the same panel and moves each price on the delivery day by $\phi$ of the announcement’s expected reaction given the reading; that part is then missing from the announcement.

| share of the market with the data | 0 | 0.25 | 0.5 | 0.75 | 0.9 |
| --- | --- | --- | --- | --- | --- |
| Sharpe ratio after costs | 4.41 | 3.20 | 1.86 | 0.45 | $-0.43$ |
| return a year after costs (gross 1) | 61.1% | 42.6% | 24.2% | 5.7% | $-5.4\%$ |
| with the data one day before the others: Sharpe ratio | 4.28 | 4.40 | 4.43 | 4.37 | 4.29 |

The value falls almost linearly with the share of the market that has the data, and turns negative before everyone has it, because the book pays ten basis points per unit traded on a one-way turnover of 36 times its gross a year. Timing is the other half. The same dataset delivered to the firm one day before the others is worth a Sharpe ratio of 4.3 to 4.4 at every level of diffusion ([Figure 16.1](#fig-s1-alternative-data-strategies-decay)): the firm trades before the others’ move and holds through it. In the dataset market this is why exclusivity and delivery speed are priced, and why vendors’ early clients earn what their later clients cannot.

![The nowcast book on the synthetic market as the alternative panel spreads: Sharpe ratio after costs against the share of the market that buys the same panel, for a firm that receives it on the same day as the others and for one that receives it a day earlier. Data: s1_altstrat.run.](https://one-course.com/images/onecourse/chapters/quant-8/s1-alternative-data-strategies/fig-b354b9b4d3b3.svg)

***Figure 16.1.** The [nowcast](#def-s1-alternative-data-strategies-nowcast) book on the synthetic market as the alternative panel spreads: Sharpe ratio after costs against the share of the market that buys the same panel, for a firm that receives it on the same day as the others and for one that receives it a day earlier. Data: `s1_altstrat.run`.*

## 16.4 Operating an alternative-data strategy

A live [alternative-data strategy](#def-s1-alternative-data-strategies-nowcast) is as much an operation as a model. The data arrive late, change format, lose coverage, and are revised; each delivery has to be checked against its own history before it moves a position (Book 7, chapter 29’s workflow). The vendor’s history must be point in time, or the backtest trades on backfilled readings fitted to the outcome (Book 7, chapter 12’s trial). The [nowcast](#def-s1-alternative-data-strategies-nowcast) needs re-estimation as the data change. And the strategy’s own decay has to be monitored: the chapter’s model suggests watching how much of the expected reaction happens before the report (the pre-report move per unit of [nowcast](#def-s1-alternative-data-strategies-nowcast)), which rises as the data spread and is observable without knowing who else buys the data.

## 16.5 Strategy files

**Strategy file 16.1 — Revenue nowcast trade.**

**Who pays you, and why.** Investors without the data, who learn the quarter’s results only at the report.

**Instruments and venues.** Stocks of companies the panel measures (retailers, consumer brands).

**Signal.** A [nowcast](#def-s1-alternative-data-strategies-nowcast) of revenue or earnings against consensus, from transactions or images.

**Sizing and execution.** Position from delivery to report, proportional to the [nowcast](#def-s1-alternative-data-strategies-nowcast); hedge the market and sector.

**Costs.** Data licences, often the largest cost; trading around reports.

**How it dies.** Diffusion of the same data to more buyers; changes in the panel’s coverage.

**Horizon, capacity, infrastructure.** Weeks; capacity limited by the covered names; a data-engineering team.

**Backtest honestly.** Point-in-time deliveries, not backfilled history; the report dates as scheduled.

**Sources.** Katona, Painter, Patatoukas and Zeng (2024); Zhu (2019); this chapter’s simulation.

**Strategy file 16.2 — Web-traffic signal.**

**Who pays you, and why.** As for [nowcasts](#def-s1-alternative-data-strategies-nowcast): online activity leads reported sales.

**Instruments and venues.** Companies whose sales happen online.

**Signal.** Visits, app usage or searches against their seasonal pattern.

**Sizing and execution.** As for [nowcasts](#def-s1-alternative-data-strategies-nowcast).

**Costs.** Data licences; noisy mapping from activity to revenue.

**How it dies.** Changes in how traffic is measured; diffusion.

**Horizon, capacity, infrastructure.** Weeks; entity mapping from domains and apps to companies.

**Backtest honestly.** Traffic as measured at the time, with the vendor’s methodology changes marked.

**Sources.** No performance figure verified; Green et al. (2019) for the related case of employee reviews.

**Strategy file 16.3 — Sentiment signal.**

**Who pays you, and why.** Investors slow to aggregate dispersed opinions about a company.

**Instruments and venues.** Stocks with enough reviews, posts or articles.

**Signal.** Changes in crowdsourced ratings or text tone (Book 7, chapter 12).

**Sizing and execution.** Monthly sorts, or tilts in a broader book.

**Costs.** Low turnover.

**How it dies.** Manipulated reviews; diffusion.

**Horizon, capacity, infrastructure.** Months; text processing.

**Backtest honestly.** Reviews timestamped when posted; deleted and edited reviews handled.

**Sources.** Green, Huang, Wen and Zhou (2019): improving employer ratings significantly outperform declining ones.

**Strategy file 16.4 — Supply-chain linkage signal.**

**Who pays you, and why.** Attention-constrained investors who do not follow news through economic links.

**Instruments and venues.** Suppliers and their principal customers.

**Signal.** The customers’ recent returns or news, mapped to their suppliers.

**Sizing and execution.** Long suppliers of customers with good news, short the others; monthly.

**Costs.** Moderate turnover.

**How it dies.** Better-known links; link data now widely available.

**Horizon, capacity, infrastructure.** A month; a supply-chain database, point in time.

**Backtest honestly.** Links as disclosed at each date; customer disclosures’ timing.

**Sources.** Cohen and Frazzini (2008): monthly alphas of over 150 basis points; Book 7, chapter 10.

## 16.6 Tutorial: by the time it is sold

**Goal.** Turn a synthetic panel into [nowcast](#def-s1-alternative-data-strategies-nowcast) positions and measure how their value falls as the data spread, and what a head start is worth. **End state:** the table and [Figure 16.1](#fig-s1-alternative-data-strategies-decay).

1. **Diffusion and the book**: others trade the data on its delivery; the book holds from delivery to report. `def diffuse (R, events, jump, phi: float , lead: int , reading, shrink: float ): """For each event (t, i) with reading m: add phi x jump[i] x shrink x m to day t - lead's return and subtract it from day t's, so the total move is unchanged and a share of it arrives when the panel is delivered.""" out = np.array(R, float , copy=True ) for (t, i), m in zip (events, reading, strict=True ): if t - lead < 0 or not np.isfinite(out[t, i]) or not np.isfinite(out[t - lead, i]): continue move = phi * jump[i] * shrink * m out[t - lead, i] = (1 + out[t - lead, i]) * np.exp(move) - 1 out[t, i] = (1 + out[t, i]) * np.exp(-move) - 1 return out def pre_event_book (events, nowcast, lead: int , T: int , N: int ): """Hold nowcast-proportional positions from the close of the delivery day (t - lead) to the close before the announcement's (weights at close d earn day d + 1), scaled each day to gross one.""" raw = np.zeros((T, N)) for (t, i), f in zip (events, nowcast, strict=True ): a = max (t - lead, 0 ) raw[a:t, i] += f g = np.abs(raw).sum(axis=1 , keepdims=True ) return np.where(g > 0 , raw / np.maximum(g, 1e-300 ), 0.0 )` **Listing 16.1.** Diffusion of the data, and the pre-event book. code/firm/altstrat/firm_altstrat.py
2. **Run** `nowcast_quality()` , `run(phi)` for five shares and `run(phi, 1)` for the head start, and `fig_altstrat.py` .

**What to change next.** Let the share with the data grow over the years and measure the book’s rolling Sharpe ratio; estimate the diffusion from the pre-report move per unit of [nowcast](#def-s1-alternative-data-strategies-nowcast); combine two panels with different delivery lags.

## 16.7 Build: alternative-data strategies

**Purpose.** [Nowcasts](#def-s1-alternative-data-strategies-nowcast) from a noisy panel, the diffusion of the same data through the market, and the pre-report book.

**Interface.** `measure(surprise, noise, rng)`, `nowcast_slope(readings, surprises)`, `diffuse(R, events, jump, phi, lead, reading, shrink)`, `pre_event_book(events, nowcast, lead, T, N)`.

**Rules.** [Nowcast](#def-s1-alternative-data-strategies-nowcast) slopes fitted on earlier events only; diffusion keeps each announcement’s total move unchanged; positions from the delivery’s close.

**Acceptance tests.** `code/firm/altstrat/tests/`: the slope of a noisy reading; diffusion moves the planted share and keeps the total; the book’s weights by hand.

**Stretch.** Diffusion growing over time; vendors with revised histories; several panels with different lags.

Sources and further reading

- C. Zhu, “Big data as a governance mechanism”, *Review of Financial Studies* 32(5), 2019.
- Z. Katona, M. O. Painter, P. N. Patatoukas and J. Zeng, “On the capital market consequences of big data: evidence from outer space”, *Journal of Financial and Quantitative Analysis* 60(2), 2025 (online 2024).
- L. Cohen and A. Frazzini, “Economic links and predictable returns”, *Journal of Finance* 63(4), 2008.
- T. C. Green, R. Huang, Q. Wen and D. Zhou, “Crowdsourced employer reviews and stock returns”, *Journal of Financial Economics* 134(1), 2019.

## 16.8 Exercises

**Exercise 16.1 ★.**

A panel reads a surprise with noise twice the surprise’s standard deviation. What is the reading’s correlation with the surprise, and the best [nowcast](#def-s1-alternative-data-strategies-nowcast) slope?

**Solution of Exercise 16.1.**

With noise twice the surprise’s spread the correlation is $1/\sqrt{1 + 2^2} = 0.45$, and the best slope is $1/(1 + 2^2) = 0.2$: the [nowcast](#def-s1-alternative-data-strategies-nowcast) shrinks the reading by four fifths. The chapter’s fitted values are 0.46 and 0.21.

**Exercise 16.2 ★.**

A stock’s daily specific volatility is 1.6% and the announcement jump is three of them per unit of surprise. With a [nowcast](#def-s1-alternative-data-strategies-nowcast) slope of 0.2 and a reading of 2, what reaction does the [nowcast](#def-s1-alternative-data-strategies-nowcast) expect, and how much of it moves on delivery if half the market has the data?

**Solution of Exercise 16.2.**

The [nowcast](#def-s1-alternative-data-strategies-nowcast) is $0.2 \times 2 = 0.4$ standard deviations of surprise, and the expected reaction $3 \times 1.6\% \times 0.4 = 1.92\%$. Half the market moves $0.5 \times 1.92\% = 0.96\%$ of it on delivery, leaving the other half for the report.

**Exercise 16.3 ★.**

Why is [data decay](#def-s1-alternative-data-strategies-decay) different from the post-publication decay of an anomaly?

**Solution of Exercise 16.3.**

Post-publication decay follows the spread of an idea that anyone can copy at no cost; [data decay](#def-s1-alternative-data-strategies-decay) follows the number of buyers of a dataset that costs money and may be sold exclusively. Its pace can be slowed by exclusivity and speed, and it can be estimated from the dataset’s sales rather than from the calendar.

**Exercise 16.4 ★★.**

Why does the book lose before the whole market has the data?

**Solution of Exercise 16.4.**

The book turns over 36 times its gross a year, one way, at ten basis points per unit traded; when most of the reaction happens on delivery, the gross return falls below the costs before the share with the data reaches one: at 0.9 the gross Sharpe ratio is 0.15 and the net $-0.43$.

**Exercise 16.5 ★★.**

Why is a one-day head start worth almost the whole strategy?

**Solution of Exercise 16.5.**

With the data a day earlier the firm is positioned before the others trade on it, so it earns their move on delivery as well as the rest at the report: the reaction’s total is unchanged, and the firm holds through all of it, whatever the share of the market that follows.

**Exercise 16.6 ★★.**

How could you estimate, from prices alone, how widely your data are used?

**Solution of Exercise 16.6.**

Measure the price move on delivery days (or in the days before reports) per unit of the [nowcast](#def-s1-alternative-data-strategies-nowcast): with no one else using the data it is zero; as usage grows it rises toward the whole expected reaction. The ratio of the pre-report move to the total estimates the share.

**Exercise 16.7 ★★★.**

*Coding.* Run `run(0.5, 0, 3.0)`, with the panel’s noise three times the surprise’s spread instead of two. What happens to the Sharpe ratio, and why?

**Solution of Exercise 16.7.**

With noise three times the spread (a correlation of 0.34) the book at half diffusion earns a Sharpe ratio of 1.35 and 17.6% a year, against 1.86 and 24.2% with noise twice the spread; without diffusion 3.20 against 4.41. A noisier panel predicts less of each surprise, so every position carries less of the reaction and the same costs.

**Exercise 16.8 ★★★.**

*Find the flaw.* “The vendor’s five-year history shows our [nowcast](#def-s1-alternative-data-strategies-nowcast) would have earned a Sharpe ratio of 3; we will pay the licence.”

**Solution of Exercise 16.8.**

A vendor’s history may be backfilled (reconstructed after the fact and fitted to the outcomes), and it does not show how many other clients now have the data. Demand point-in-time deliveries, trial the data live for a period, and estimate its current diffusion from the pre-report moves.

## 16.9 Problem: By the Time It Is Sold

**Problem 16.1.**

Weekend problem — a dataset’s life

The chapter’s synthetic panel and the public record.

**Part I — From data to position.**

1. Define a [nowcast](#def-s1-alternative-data-strategies-nowcast) and an [alternative-data strategy](#def-s1-alternative-data-strategies-nowcast) .
2. List the four steps from a dataset to a position.
3. Describe the chapter’s panel and its [nowcast](#def-s1-alternative-data-strategies-nowcast) .
4. What does the book hold, and when?

**Part II — The record.**

5. What did Katona and co-authors find about satellite data?
6. What did Green and co-authors find about employer reviews?
7. What did Cohen and Frazzini find about economic links?
8. What did Zhu find about price informativeness?

**Part III — Decay.**

9. Define [data decay](#def-s1-alternative-data-strategies-decay) and describe how the chapter plants it.
10. Give the Sharpe ratios as the share with the data grows.
11. Why does the value fall almost linearly?
12. What does a head start do?

**Part IV — The verdict.**

13. State the *named result* : the [nowcast](#def-s1-alternative-data-strategies-nowcast) strategy’s Sharpe ratio as the share of informed capital using the data grows.
14. How would you monitor a dataset’s decay?
15. What operational risks does the strategy carry?
16. How would you evaluate a vendor’s history?
17. What should a data licence be worth to the firm?
18. Which strategy file decays fastest?
19. How do these strategies relate to chapter 7?
20. In one sentence: what does an alternative dataset sell?

**Solution of Problem 16.1.**

1. An estimate of a later-reported quantity from earlier data; a strategy trading on such [nowcasts](#def-s1-alternative-data-strategies-nowcast) from non-traditional data.
2. Map the data to companies, [nowcast](#def-s1-alternative-data-strategies-nowcast) the reported quantity, translate it into an expected reaction, and hold over the window in which the market learns.
3. A tenth of the companies, readings with noise twice the surprise’s spread (correlation 0.46), delivered ten days early; slope 0.21 from the first two years.
4. Nowcast-proportional positions from the delivery’s close to the announcement’s close, gross one.
5. Satellite data let sophisticated investors target bad reports, with more informed shorting and lower liquidity: more asymmetry, not immediately better prices.
6. Improving employer ratings outperform declining ones and forecast earnings surprises.
7. Prices ignore news about principal customers; a supplier strategy earned over 150 basis points a month.
8. Alternative data raised price informativeness by lowering information costs.
9. The loss of value as more investors trade the same data earlier; a share $\phi$ moves $\phi$ of the expected reaction on delivery.
10. 4.41, 3.20, 1.86, 0.45 and $-0.43$ for shares of 0, 0.25, 0.5, 0.75 and 0.9.
11. Each buyer moves an equal part of the reaction to the delivery day, out of the book’s reach.
12. It keeps the Sharpe ratio near 4.3–4.4 at every level of diffusion.
13. **Named result.** The [nowcast](#def-s1-alternative-data-strategies-nowcast) book’s Sharpe ratio after costs falls from 4.41 when no one else has the panel to 1.86 when half the market has it and $-0.43$ at nine tenths, while a firm that receives the panel a day earlier keeps 4.28 to 4.43.
14. Track the pre-report move per unit of [nowcast](#def-s1-alternative-data-strategies-nowcast) and the book’s realised reaction per trade.
15. Late or changed deliveries, coverage losses, revisions, vendor failures and licence terms.
16. Insist on point-in-time history, check for backfill, and trial it live.
17. What it adds to the firm’s returns after costs over its expected life, given its diffusion.
18. The revenue [nowcast](#def-s1-alternative-data-strategies-nowcast) trade, whose data are sold to many buyers.
19. They trade the same reports before they happen; chapter 7 trades them after.
20. Time: knowledge of a quantity before the market has it.

## 16.10 Interview questions

**Interview question 16.1 ★ researcher.**

How would you evaluate a new alternative dataset?

**Solution of Interview question 16.1.**

Coverage and history (point in time?), the mapping to companies, the correlation with the reported quantity on data the vendor could not have fitted, the incremental IC over the firm’s existing signals, the lag, the cost, who else buys it, and the legal terms.

**Interview question 16.2 ★★ researcher.**

What is backfill bias in vendor data, and how do you detect it?

**Solution of Interview question 16.2.**

History reconstructed after the fact, often fitted to known outcomes, so the backtest looks better than live use. Detect it by comparing fit before and after the vendor’s launch date, by coverage jumps and level shifts at the launch, and by asking for delivery timestamps (Book 7, chapter 12).

**Interview question 16.3 ★★ researcher, trader.**

Your alternative-data signal has weakened over two years. What might be happening?

**Solution of Interview question 16.3.**

More buyers (diffusion), a change in the panel (provider, coverage, methodology), a change in what the companies report, or crowding by similar datasets. Check the pre-report move per unit of signal and the panel’s own statistics.

**Interview question 16.4 ★★ developer.**

Design the pipeline that takes a daily vendor delivery to a trade the same evening.

**Solution of Interview question 16.4.**

Ingest and validate the delivery (schema, coverage, ranges against history), map to company identifiers point in time, compute [nowcasts](#def-s1-alternative-data-strategies-nowcast) with the current model, pass them to the optimiser with the other signals, generate orders for the close, and log everything for reproduction.

**Interview question 16.5 ★★ risk.**

What are the legal and compliance risks of alternative data?

**Solution of Interview question 16.5.**

Material non-public information (data obtained in breach of a duty), personal data and privacy law, licence terms and exclusivity, and web-scraping terms of service; every dataset needs a compliance review of how it was collected.

**Interview question 16.6 ★★★ researcher.**

A share $\phi$ of the market trades a signal on its delivery and moves the price by $\phi$ of the expected reaction. Derive the expected P&L of a trader who receives it at delivery, and of one who receives it $k$ days earlier and holds through.

**Solution of Interview question 16.6.**

With expected reaction $J$ per unit of [nowcast](#def-s1-alternative-data-strategies-nowcast), a trader at delivery earns $(1 - \phi)J$ per unit, since $\phi J$ moves when the others trade. A trader with the data $k$ days earlier, holding through, earns the whole $J$ (the others’ move included), less any drift of the extra days; the value of speed is $\phi J$.