Alternative Data for Asset Management
Question: how does an observation about economic activity become a defensible input to an investment decision?
A card transaction, a paragraph in a filing, and a link between companies measure different things. Before fitting a model, we need to understand how each observation was collected, which entities it represents, and when an investor could have used it. More records do not automatically produce a better measure of the underlying economic activity.
Course status: this overview and preparation guide establish the alternative-data sequence. Chapters 4.1–4.4 are planned. Begin with the Machine Learning preparation guide for features, targets, and chronological validation.
Before You Begin
You should be able to group and join pandas tables, work with timestamps and missing values, and explain the difference between model fitting and out-of-sample evaluation. The planned lessons will introduce text and network methods through small examples before larger datasets.
Keep three questions from the existing portfolio lessons in view:
- What is being measured? An observed proxy and an economic quantity can differ, just as factor exposures and actual holdings differ in Chapter 2.1.
- What was known at the decision date? Apply the information timeline in the Course 3 overview.
- How does this affect a portfolio? Apply common constraints and implementation costs as in the capstone.
The Planned Learning Sequence
4.1 — Consumption
Examine transaction, location, and activity records as proxies for consumption. Aggregate them by time and entity, investigate seasonality and changing sample coverage, and formulate a specific economic hypothesis. Distinguish a change in activity from a change in the panel that records it.
Experiment: compare raw activity with a coverage-adjusted measure, then change the aggregation window. ECharts time series and seasonal heatmaps will make the measurement choices visible; geographic displays will be used where location contributes to the question.
4.2 — Textual Analysis for Financial Applications
Turn a small collection of documents into numerical features. Work through tokenization, stop words, term counts, TF-IDF, and similarity. Explain why the vocabulary and document-frequency statistics are fitted on the training corpus when the task is predictive.
Experiment: change a preprocessing choice and inspect term-weight bars and document-similarity heatmaps. Removing a negation such as “not” can change the financial meaning, so preprocessing choices require interpretation rather than automatic acceptance.
4.3 — Processing Corporate Filings
Develop two connected analyses:
- 10-K reports: compare business and risk language, document similarity, and sentiment. Separate substantive changes from repeated or boilerplate language.
- 13-F reports: compare reported holdings, portfolio overlap, and institutional similarity. Account for the scope of reported securities and the delay between the holdings date and public availability.
Experiment: select two filings or institutions and examine the source observations alongside their similarity measures. Charts will distinguish the reporting period from the date the information became available.
4.4 — Using Media-Derived Data
Aggregate media or social observations into dated sentiment and relationship features. Construct company networks, explain degree and PageRank with a small graph, and investigate how sources, time windows, and edge definitions affect the result.
Experiment: adjust the observation window or network threshold and compare sentiment series and network structure. A highly connected company is not necessarily a more attractive investment; the graph measures the relationship defined by its edges.
Define the Observation Before the Feature
For each dataset, write a short data dictionary before calculating a score:
| Field | Question it answers |
|---|---|
| Entity identifier | Which company, location, document, or institution does the row describe? |
| Observation period | When did the underlying activity occur, or what period does the filing report? |
| Available-at timestamp | When could the chosen data source have supplied this information? |
| Value and units | Is this a count, currency amount, location, text field, or portfolio position? |
| Coverage and missingness | What does absence mean, and which entities are represented? |
| Source and snapshot | Which release or version produced this reproducible record? |
Joins require particular care. A company name is not a stable identifier, and today's entity mapping may include information unavailable in an earlier historical period. A zero transaction count and a missing observation must not silently become the same measurement.
Work Through Two Measurement Decisions
Activity Versus Panel Growth
Suppose a transaction panel records 1,000 purchases from 100 observed customers in one month and 1,200 purchases from 150 in the next. Raw purchases rise by 20%, but purchases per observed customer fall from 10 to 8, a decline of 20%.
Neither calculation alone establishes total consumer demand. The customers may have changed, their coverage may be uneven, and the panel may not represent the population. A useful next step is to compare a common customer cohort and report how much of the original panel remains.
Check your reasoning: which comparison answers “did recorded purchases increase?” Which answers “did the average observed customer purchase more?” Explain why both can be correct while pointing in different directions.
A Filing's Period Is Not Its Availability Date
Suppose a holdings report describes positions at 30 June but becomes public after market close on 14 August. A backtest using that report cannot allocate on 30 June. It must use a decision after publication, with a stated processing and trading delay—for example, the following trading day's close if that is the chosen convention.
Record amendments as later versions rather than silently replacing the original history. Likewise, a sentiment score extracted from a later filing should not be assigned to an earlier trading date merely because the filing discusses an earlier reporting period.
From a Feature to an Investment Question
Write the chain explicitly:
Source observation → cleaning and aggregation → dated feature → prediction or comparison → investment decision → evaluation.
At each step, state the assumption. For example, a negative-word fraction measures the frequency of a vocabulary, not an investor's loss probability. A holdings-overlap score measures similarity within reported positions, not necessarily similarity of complete portfolios. Establish a baseline and inspect examples before attaching a financial interpretation.
The lessons will use reproducible, documented data snapshots served with the website. Small worked examples, editable Python, ECharts, and selective ipywidgets experiments will let students inspect how the feature changes instead of treating a score as a black box.
Syllabus Reference
This sequence follows the four modules in EDHEC's Python and Machine-Learning for Asset Management with Alternative Data Sets: Consumption, Textual Analysis for Financial Applications, Processing Corporate Filings, and Using Media-Derived Data.