Trading platform, financial marketsInternal

Market research straight from the raw transaction record

At a trading platform, AI writes the scripts that pull trades and orders from a public ledger, then helps turn the raw records into product and market decisions.

Industry: Trading platform, financial markets
What it shows: Data extraction, Market analysis, AI-assisted research
Year: 2026
Status: Internal

Illustration. Markets, figures and findings are invented; the platform stays unnamed.

What changes

The steps people did by hand, and the route the work takes now.

The team needs to know how a market really behaves

Before

Wait for a data request
Export what the dashboard happens to show
Clean it in a spreadsheet
Guess at the rest

With AI

Claude writes and reruns the extraction scripts
Clean tables are analysed and checked against ledger totals

Product decisions backed by the full trading record

The problem

A trading platform makes product calls every week: which markets to deepen, how to route orders, where new listings need support. The answers were sitting in the raw trading record, but that record is enormous, low level and spread across millions of events, and the usual dashboards only show what someone thought to build in advance.

What we built

A research workflow where AI does the heavy lifting on data that would normally need a data engineering team. Claude writes the scripts that pull trades, orders and balance changes straight from the public ledger, turns them into clean tables and checks them against the ledger’s own totals. The analysis runs on the full record rather than a sample, and every finding ends in a decision for the product team.

Why it matters

Questions that used to wait in a queue for a data request are answered from the source in a day. Because the scripts are kept and rerun, a follow-up question costs minutes, and every figure in a research note traces back to raw records.

Inside the research

Pick a question and run it from script to decision, browse the markets behind it, or see the method that keeps every figure tied to the ledger.

Interactive · 6 journeys

Interactive system map

From a question to the ledger and back as a decision

12 parts and 6 journeys: the reviewed script, the full record, the check against the ledger's totals and the research note. The numbers below are process facts, not platform figures.

Play 6 journeys, or click any of its 12 parts.

stages, from the public ledger to a decision: 5
rules every figure follows: 4
example questions, each run from script to decision: 3

Explore the map

InteractivePick a question and run the research, or sort and filter the markets.

Ask the ledgerMarketsMethod

Pick a research question.

Question 1Where volume concentratesQuestion 2Small orders on thin marketsQuestion 3How long new markets last

The question

Which markets carry the trading, and how deep are their order books?

ExtractClaude writes the scriptWaitingReading the ledgerDone

extract_trades.pyv3Written with Claude

# Written with Claude, reviewed, kept under version control
from ledger_client import PublicLedger
import pandas as pd

ledger = PublicLedger()
rows = []
for market in ledger.markets():
    for t in ledger.trades(market, since="2026-06-01"):
        rows.append({"time": t.time, "market": market,
                     "side": t.side, "size": t.size,
                     "price": t.price})

trades = pd.DataFrame(rows)
trades["value"] = trades["size"] * trades["price"]
share = trades.groupby("market")["value"].sum()
share = 100 * share / share.sum()

Claude drafts the script from the question and a person reviews it before it runs. It reads every record in the period, not a sample.

Clean tableReady for analysisWaitingCleaningDone
trades, first 5 of 1,284,906 rows
Preview of the clean trades table, dummy rows
time market side size price
00:00:04 A buy 1,250 4.0821
00:00:04 A sell 300 4.0819
00:00:11 B buy 72 0.5530
00:00:19 D sell 18 12.410
00:00:26 C buy 5,000 0.0914
Raw records become a tidy table with consistent types, units and market labels, so the analysis starts from the same ground every time.
3
Check against the ledgerBefore any analysisWaitingComparing totalsDone
1. TradesExtracted 1,284,906Ledger 1,284,906Match
2. VolumeExtracted 9,402,118,330Ledger 9,402,118,330Match
Analysis only starts once every line matches the ledger’s own totals.
Row counts and totals are compared with the ledger’s own figures. A mismatch stops the work until it is explained.
4
Analyse and decideFinding and decisionWaitingAnalysingDone
Share of volume by market
Last 90 days, all trades
Market A31%, deep order book
Market B19%, deep order book
Market C12%, deep order book
Market D9%, thin order book
Market E7%, thin order book
Market F5%, thin order book
Market G4%, thin order book
Other markets13%
Finding
Three markets carry 62% of the volume. The thin markets below them trade often, but in small amounts and at wide spreads.
Decision for the product team
Put depth and routing work into Markets A to C first.
Every figure here traces back to raw ledger records through extract_trades.py, v3.

Preview of the clean trades table, dummy rows
time	market	side	size	price
00:00:04	A	buy	1,250	4.0821
00:00:04	A	sell	300	4.0819
00:00:11	B	buy	72	0.5530
00:00:19	D	sell	18	12.410
00:00:26	C	buy	5,000	0.0914

The question

Do small orders fill on thin markets as often as on deep ones?

ExtractClaude writes the scriptWaitingReading the ledgerDone

extract_orders.pyv2Written with Claude

# Written with Claude, reviewed, kept under version control
from ledger_client import PublicLedger
import pandas as pd

ledger = PublicLedger()
rows = []
for market in ledger.markets():
    for o in ledger.orders(market, since="2026-06-01"):
        rows.append({"order": o.id, "market": market,
                     "size": o.size, "filled": o.filled_size,
                     "hours_open": o.hours_open})

orders = pd.DataFrame(rows)
orders["full"] = (orders["filled"] == orders["size"]) \
    & (orders["hours_open"] <= 1)
bands = pd.cut(orders["size"], [0, 100, 1_000, 10_000, None])

Claude drafts the script from the question and a person reviews it before it runs. It reads every record in the period, not a sample.

Clean tableReady for analysisWaitingCleaningDone
orders, first 5 of 3,906,442 rows
Preview of the clean orders table, dummy rows
order market size filled full
o-118204 A 40 40 yes
o-118205 F 25 0 no
o-118207 B 2,400 2,400 yes
o-118211 G 60 35 no
o-118216 D 15,000 15,000 yes
Raw records become a tidy table with consistent types, units and market labels, so the analysis starts from the same ground every time.
3
Check against the ledgerBefore any analysisWaitingComparing totalsDone
1. OrdersExtracted 3,906,442Ledger 3,906,442Match
2. Filled sizeExtracted 61,207,559Ledger 61,207,559Match
Analysis only starts once every line matches the ledger’s own totals.
Row counts and totals are compared with the ledger’s own figures. A mismatch stops the work until it is explained.
4
Analyse and decideFinding and decisionWaitingAnalysingDone
Orders filled in full within an hour
By order size, in units
Under 100 units: deep 97%, thin 64%
100 to 1k units: deep 98%, thin 79%
1k to 10k units: deep 99%, thin 90%
Over 10k units: deep 99%, thin 95%
Finding
On thin markets, orders under 100 units fill in full only 64% of the time, against 97% on deep markets. The gap closes as orders grow, because larger orders mostly cross the spread.
Decision for the product team
Route small orders on thin markets through the deeper pool by default.
Every figure here traces back to raw ledger records through extract_orders.py, v2.

Preview of the clean orders table, dummy rows
order	market	size	filled	full
o-118204	A	40	40	yes
o-118205	F	25	0	no
o-118207	B	2,400	2,400	yes
o-118211	G	60	35	no
o-118216	D	15,000	15,000	yes

The question

How long do new markets keep the activity they have at launch?

ExtractClaude writes the scriptWaitingReading the ledgerDone

extract_listings.pyv4Written with Claude

# Written with Claude, reviewed, kept under version control
from ledger_client import PublicLedger
import pandas as pd

ledger = PublicLedger()
rows = []
for market in ledger.markets(listed_since="2026-01-01"):
    for t in ledger.trades(market, since=market.listed):
        rows.append({"market": market, "account": t.account,
                     "week": t.weeks_since_listing + 1})

df = pd.DataFrame(rows)
weekly = df.groupby(["market", "week"]).agg(
    trades=("account", "size"),
    accounts=("account", "nunique"))

Claude drafts the script from the question and a person reviews it before it runs. It reads every record in the period, not a sample.

Clean tableReady for analysisWaitingCleaningDone
listing_weeks, first 5 of 72 rows
Preview of the clean listing_weeks table, dummy rows
market listed week trades accounts
K 07-06 1 9,412 1,206
K 07-06 2 4,630 702
K 07-06 3 4,071 611
L 07-20 1 7,958 1,044
L 07-20 2 4,188 598
Raw records become a tidy table with consistent types, units and market labels, so the analysis starts from the same ground every time.
3
Check against the ledgerBefore any analysisWaitingComparing totalsDone
1. TradesExtracted 418,230Ledger 418,230Match
2. AccountsExtracted 26,915Ledger 26,915Match
Analysis only starts once every line matches the ledger’s own totals.
Row counts and totals are compared with the ledger’s own figures. A mismatch stops the work until it is explained.
Analyse and decideFinding and decisionWaitingAnalysingDone
Activity after listing
Nine new markets, week one = 100
Trades by week since listing: 100, 52, 44, 41, 39, 38, 38, 37. Active accounts: 100, 61, 54, 50, 48, 47, 46, 46.
Finding
Activity in a new market roughly halves in week two, then settles near 40% of launch week. Active accounts fall less steeply than trades.
Decision for the product team
Extend launch support through the second week, and review each new market at week four.
Every figure here traces back to raw ledger records through extract_listings.py, v4.

Preview of the clean listing_weeks table, dummy rows
market	listed	week	trades	accounts
K	07-06	1	9,412	1,206
K	07-06	2	4,630	702
K	07-06	3	4,071	611
L	07-20	1	7,958	1,044
L	07-20	2	4,188	598

Order book depth

All12Deep5Thin7

12 markets, 100% of volume. Deep markets have a median spread of 6 bps, thin ones 42 bps.

5 deep markets carry 69% of volume, at a median spread of 6 bps.

7 thin markets carry 31% of volume, at a median spread of 42 bps.

Twelve dummy markets with volume share, order book depth, median spread, small-order fill rate and 30-day trend. Sortable.
	Depth
Market Adeep	deep	31%	3 bps	98%	+6%
Market Bdeep	deep	19%	4 bps	97%	+2%
Market Cdeep	deep	12%	6 bps	97%	−3%
Market Dthin	thin	9%	24 bps	71%	+9%
Market Ethin	thin	7%	31 bps	66%	−4%
Market Fthin	thin	5%	38 bps	63%	0%
Market Gthin	thin	4%	45 bps	60%	−8%
Market Hdeep	deep	4%	9 bps	96%	+1%
Market Ithin	thin	3%	42 bps	64%	−2%
Market Jdeep	deep	3%	11 bps	97%	+4%
Market KNewthin	thin	2%	61 bps	58%	−38%
Market LNewthin	thin	1%	55 bps	66%	−46%

Volume share over the last 90 days. Spread is the median, in basis points. Small-order fill is the share of orders under 100 units filled in full within an hour. The trend is the change in daily trades over 30 days, on one scale for every market.

01SourceEvery trade, order and balance change in the public ledger. The full record, not a sample.Out: Raw ledger records
02ExtractClaude writes the extraction scripts against the ledger’s public interfaces.Out: Versioned scripts
03Clean and checkOne row per event, tied to the ledger’s own totals before anyone looks for a pattern.Out: Clean tables
04AnalyseLiquidity, trading activity and how participants behave, market by market.Out: Charts and figures
05DecideEach note ends in a recommendation the product team can act on.Out: A decision

What the method guarantees

Full record, not a sampleScripts read every event in the period, so small markets and rare cases are counted too.
Scripts versioned and rerunWhen a question changes, the script changes and runs again from the first record.
Tied to ledger totalsRow counts and totals must match the ledger before any analysis starts.
Traceable figuresEvery figure in a research note points back to the raw records it came from.

One script, four versions

extract_listings.py, behind the question on new markets. Each change reran it from the first record.

v4Count active accounts per week, not only trades2026-08-31Rerun, totals match
v3Start each market from its own listing date2026-08-12Rerun, totals match
v2Add the markets listed in July2026-07-28Rerun, totals match
v1First version: trades per new market per week2026-07-09Rerun, totals match

Illustration. Markets, figures and findings are invented; the platform stays unnamed.

What it shows

Most companies sit on a raw record they never fully use: transaction logs, event streams, ledgers, order histories. AI is now good enough to write and maintain the extraction code as well as read the results, which puts that record within reach of a small team.

Under the hood

Claude
Python
SQL

01Source
Every trade, order and balance change sits in a public transaction ledger. Nothing is sampled; the full record is the starting point.
02Extract
Claude writes the extraction scripts against the ledger's public interfaces. They are versioned and rerun from the start when a question changes.
03Clean and check
Raw records become one-row-per-event tables, checked against the ledger's own totals before anyone draws a conclusion.
04Analyse
Liquidity, trading activity and how participants behave, market by market, with AI helping to explore and to explain what the numbers show.
05Decide
Each piece of research ends in a recommendation the product team can act on, with the figures traceable back to raw records.

What changes

The problem

What we built

Why it matters

Inside the research

From a question to the ledger and back as a decision

Which markets carry the trading, and how deep are their order books?

Do small orders fill on thin markets as often as on deep ones?

How long do new markets keep the activity they have at launch?

What the method guarantees

One script, four versions

What it shows

Under the hood

Want something like this built for your team?