Capture.

Ten sources, nine asset classes, one store

A mid-frequency system needs long daily history to test across many market conditions, and dense intraday bars to measure entries, exits and costs. Nexus holds both in one time-series store, so research, the simulator and the trader read from the same place.

What can be tradedExecution today is through one broker, IG, on a demo account. Of its 64 captured markets, 37 are enabled for the live book and 27 are captured for research only. Everything else on this page is research data.

From expiring contracts to one continuous series

A trend book needs long, unbroken futures histories, but a futures contract expires every month or quarter. A continuous series joins the contracts together. Each choice in that join changes the backtest: which contract, when to switch, and how to close the price gap at the switch. Here is each choice Nexus makes.

1. 1

   Which contract

   Front and next, hour by hour

   For each of 40 markets the exchange feed supplies two hourly series: the front contract and the contract after it. It also supplies every individual contract daily, 3,982 of them since 2010. The front contract is chosen by the vendor's published rule, not by hand.

   Runs

   Hourly and daily, for all 40 markets.

2. 2

   When to switch

   Volume for most, the calendar for rates

   35 markets switch when the next contract trades more volume than the front. The five rates markets, four Treasury futures and three-month SOFR, switch on the calendar, the day after the front contract's last trade. In September 2026 the 10-year Treasury last traded on the 21st and switched on the 22nd. The E-mini S&P 500 switched on volume on the 16th, two days before its last trade.

   Nexus also keeps its own roll calendar for 47 markets: 8,564 contracts from 1999 to 2031, each rolled five business days before its last trade. For Treasuries the calendar rolls five business days before an approximate first-notice date instead. It decides which contract the term-structure inputs use. It does not set the continuous price series.

   Runs

   The vendor's rule sets the price series. The house calendar skips weekends but not holidays.

3. 3

   How to close the gap

   Ratio back-adjustment

   At a switch the next contract usually trades at a different price from the front, and that gap is not a return anyone earned. Nexus removes it by scaling older history by the ratio of the two prices at the switch, as the next section shows.

   Runs

   On all 40 exchange markets.

4. 4

   Where the live book stands

   The broker's contracts

   The book trades the broker's contracts, not the exchange's. The broker's undated contracts never expire. Its dated ones, such as the Treasury contracts, move to the next contract around the last dealing date. When they do, the live series is re-anchored so the switch adds no return (section 06).

   Runs

   On the broker demo account.

Ratio by default, difference where it fits

No way of joining contracts is neutral. Each method keeps one property and gives up another. Nexus picks a method per source, records it on every dataset, and checks that the result still agrees with the raw prices.

• The ratio

  A roll is wherever the front contract changes. The ratio is the next contract's close over the front contract's close on the last bar before the switch, and all earlier history is scaled by the product of the ratios after it. Where the next contract has no price at that bar, the ratio falls back to the new front contract's first price over the old one's last. Those prices come from different bars, so the market's move between them is folded into the adjustment instead of kept as a return. A ratio outside 0.5 to 2 is flagged, and one outside 0.1 to 10 is not applied.

• Negative prices

  US crude oil settled at −$37.63 on 20 April 2020. A ratio cannot span a change of sign, so negative bars are left out of the ratio-adjusted series, counted and warned about rather than dropped in silence. A difference-adjusted variant for those markets is not built.

• New data, old history

  Each new window of exchange data is adjusted on its own, then scaled to the stored series at the latest bar both share. A new roll therefore never rewrites history. A join that moves more than 0.15 in log terms is refused and alerted, never written. Between full rebuilds, recent adjusted prices drift away from the traded contract price, while returns are unaffected.

• Adjustments on record

  Each market's dataset records the adjustment events found in its history, each with the time it takes effect and its ratio: 4,229 across 40 markets. They are detected afterwards from the steps between the adjusted and raw series, and refreshed in batches. Only 71 name the two contracts. A very small adjustment can fall below the detector's tolerance, and the recorded time is when the adjustment applies, not necessarily the vendor's roll date. They are not a complete roll ledger.

• Checked against raw

  The data checks compare adjusted with raw. Three can fail a series: the latest adjusted level must anchor to the raw price, the adjustment must not crush the size of price moves, and the declared method must suit the data. Two more only warn: roll returns well beyond the market's own, and long-run drift between adjusted and raw. Details are in Validate and clean.

• Roll cost

  The adjusted series removes the gap but not the cost of rolling. On markets traded through dated contracts, the simulation charges a smoothed roll cost of four spreads a year.

For a trend bookA trend book sizes positions from percentage returns and their volatility. Ratio adjustment keeps those true across a roll where both contracts are quoted at the switch, which is why it is the default. Point-value history is still there in the raw contracts kept beside it.

Every series side by side, none overwritten

Nexus keeps the raw and the derived together. Each dataset is tagged with its source, how it was built, its roll method and rule, its calendar, its time zone and how its prices are quoted. Any adjusted price can be traced back to what the vendor delivered.

And the clean layerThe cleaning engine does not edit these series. A clean copy is rebuilt from raw plus a log of operations, and every operation can be undone. That is step 2, Validate and clean.

Broker ticks, stamped on arrival and written once

The broker is both a data source and the venue. Capturing its own prices means the live book and its record see exactly what the broker quoted.

• Stamped on arrival Runs

  Every tick is stamped with Nexus's own UTC receipt time, because the broker's update time has no date. Bid, offer and mid are all kept.

• Written once Runs

  A bar holds only the ticks inside its own interval and is finalised once. A tick that arrives late is dropped, never applied to a closed bar. 1-second and 1-minute bars go to the time-series store. A 1-second quote cache serves the trader, but it is not the record.

• Gaps stay gaps Runs

  A second or a minute with no ticks stays empty, and no bar is ever made up. Each weekday night, gaps of 30 minutes or more from the last 14 days are fetched again from the broker's own price history. Which minutes should exist is learned from each market's last four weeks: a minute is expected if it appeared in at least 60 % of them. The refetch stays inside a fixed share of the broker's weekly data allowance.

• Derived bars Runs

  5, 15 and 60-minute bars are rebuilt from the 1-minute bars every ten minutes. Only fully elapsed buckets are written, so a half-formed hour never reaches research.

• Watchdogs Runs

  The capture reconnects and alerts after 120 seconds without a connection, and after 300 seconds with no finished 1-minute bar while a market is open. A subscription that falls silent is renewed every 180 seconds, up to twice, before its connection is rebuilt, at most six times an hour. A sweep each morning pages operations about any hour with minute bars but no hourly bar. At 12:34 UTC on 15 October, 88 of 91 subscriptions were receiving.

Under the hoodThe broker's streaming protocol, and why a connection can report healthy while writing nothing, are in Three ways to stream.

History and the live tail, joined at a seam you can inspect

Exchange history arrives twice a day, hours after the close, while the broker's prices arrive every second. Simulation needs the long history and a live book needs the latest hour. Nexus joins the two into one series, and simulation and the live book both read it.

1. 1

   A byte copy up to the handoff

   Up to the last completed bar that both sources share, the blended series is an exact copy of the ratio-adjusted exchange series. The build checks that this prefix matches byte for byte, so the history a strategy is tested on is the exchange's.

   Runs

   Rebuilt hourly, with a full recompute every night as a check.

2. 2

   Broker returns after it

   From the handoff on, the broker's mid-price returns are chained onto the last exchange level, with open, high and low scaled by the same factor. The handoff moves on every hourly run, so each exchange delivery replaces the broker's prices with the exchange's. Exchange data lands twice a day, about 14 hours after the bar at the median, so the broker tail is typically 4 to 13 hours long.

   On dated broker contracts the broker and exchange prices differ by about zero. On the broker's undated contracts the gap is the funding rate. Outside grains, softs and energy, the error at the join is within 3.4 basis points.

   Runs

   For 36 markets.

3. 3

   Broker rolls add no return

   When the broker moves a dated contract to the next month, its price steps by the calendar spread. Nobody made or lost that money. At the 27 August 2026 rates roll this step would have put four false returns, of 0.06 % to 0.38 %, into the series the live book trades from, each far too small for an outlier check to catch. The series is re-anchored at every broker roll so the seam return is exactly zero.

   Runs

   The bar at the seam loses its true move along with the false one.

4. 4

   Unexplained steps are refused

   A first broker-tail return beyond 0.05 in log terms, about 5 %, is refused, not written. A market refused six runs in a row is reported stale. The tail's age raises a warning at 36 hours of open market, not 36 hours of clock time, so a weekend is not an alarm.

   Runs

   A refused market is not updated on that run, and the refusal is logged.

For your firmA strategy is tested on exchange history and traded on the broker's prices, and the join between them is a rule you can read. The seam, its guards and its known limits are all on record.

What runs, stated plainly

Naming what is not built yet, and what is still being checked, is part of the method.

• This page describes the data platform. It shows no strategy returns and makes no performance claim.
• Figures move as data lands; each is dated to the day it was read.

Sources, read 2026-10-15: TimescaleDB catalog estimates and compression statistics; the store's coverage cache; the dataset, roll-calendar, contract and calendar reference tables; the logged cleaning and roll records; the data service's fetch, adjustment and splice code on its main branch; the capture service's status; and the work board for the open investigations.