ALMAZOV TRADE
ALMAZOV TRADE
AccountDocumentationContact the managerEN/RU
← Research & Development

Data engineering

Historical data and the state of a trading algorithm

Historical testing depends on more than the entry rule. Data continuity, time aggregation and the state accumulated before a decision can all change what the algorithm sees.

Preparing a usable history

In March 2026, our development work obtained Binance data for testing and prepared a volume view in Grafana. Later discussions dealt with gaps between historical segments and the need to rebuild a continuous history before launch.

A gap is not just an empty patch on a chart. For a model that accumulates state, the sequence arriving before the next event contributes to its initial conditions. A visual comparison alone cannot establish that two runs had equivalent inputs.

Shared processing, explicit initial conditions

In one system described in June 2026, historical and incoming data were processed by the same engine. The team also discussed how much history was needed to initialise the model and how historical processing affected its state.

Sharing the processing logic reduces the number of separate implementations that must be reconciled. It does not by itself prove deterministic replay, the absence of future information or uninterrupted operation of every current system.

Research architecture

What has to stay consistent across a run

The same market is not always the same input

Instrument, venue, interval and event timestamps define the input. A change in any of them can change the sequence of decisions.

  • Identify the data source and instrument
  • Preserve the order of events
  • Distinguish a data gap from an absence of trading
Continuity before a decision

Missing segments affect both the sample and accumulated state. Rebuilding the history is part of preparing a comparable run.

  • Record missing segments
  • Keep the aggregation rule explicit
  • Preserve the boundaries of the available history
Initialisation is part of the model

Two runs with the same last candle may have different preceding states. The development history includes shared processing of historical and incoming data.

  • Identify the initial state
  • Record how history contributes to initialisation
  • Compare intermediate state, not only final PnL
A decision must become an admissible order

The model’s intended position meets contract steps, remaining size and execution costs. These assumptions belong to the run.

  • Separate requested and executable quantity
  • Update remaining size after a fill
  • Identify which fees and slippage assumptions are included
Trace a difference back to its cause

A discrepancy in the final result can begin in the data, a rule or execution. Comparable inputs and intermediate decisions make the investigation specific.

  • Identify the version of the rules
  • Keep the sequence of events and decisions
  • Separate an engineering correction from proof of trading performance

From processed data to useful analytics

The current volume analytics cover BTC and ETH at Local and Global scales. Each buying and selling line has its own inner sum and outer sum. The displayed percentage describes activity relative to that side’s baseline.

Grafana displays the relative volumes; a separate proprietary algorithm selects impulses for AlmazovTrade Support notifications. The owner describes its development as based on backtests of six years of market data. The dated captures below show a single BTC interval at two scales, retaining their original axes and values.

Original Grafana captures

One BTC interval, two scales

Local

BTC Local: original relative-volume chart. The main sell peak is about 22.9%, with buys at about 12.1% at the same point.
–

Global

BTC Global: original chart for the same period as BTC Local. Sells peak at about 1.84%, with buys at about 1.78% at that point; readings remain below 6%.
–

Both views cover the same period. Read the numerical percentages: their vertical scales differ.