Skip to content

Replay a live session from a checkpoint

Start a backtest inside a live session's state, so it matches the session and can test another configuration from the same point. Without one, a backtest over the session's journal lacks the session's positions, balances and learnt strategy state, and parts from the session at its first order. Why each step is there: Backtesting.

Record checkpoints

Add to the session's configuration:

[engine]
checkpoint_s = 300

Every 300 s of engine time the journal gets a checkpoint: the strategy's state() if it keeps state, the positions and the open orders (Journal format). Besides the market, a backtest also reads back from the journal:

  • the session's own orders, with their send, venue and receive times;
  • the balances its venues reported;
  • money moved in or out of its accounts outside the engine (Balances);
  • the parameter changes it received.

Backtest from a checkpoint

Copy the configuration the session ran with and add:

[backtest]
fill_model = "l2_queue"
md_arrival = "recorded"
initial_state = "journal"
start_at = "2026-10-20T08:00:00Z"
params_from_journal = true
balances_from_journal = true
latency_from_journal = true

Run it over the journal from a part well before start_at. The run reads the data from that part on and starts trading at the last checkpoint at or before start_at.

./build/release/bin/fastmm-backtest --config replay.toml --data "journal:runs/mm.12.fmj,strip_own=1" --journal-out runs/replay/session.fmj --out runs/replay
  • The log names the checkpoint it started from (journal: started from the checkpoint of <ns> ns); a start_at before the first checkpoint stops the run with exit code 3.
  • Give a part early enough for the books and the strategy's averages to settle before the checkpoint.
  • Keys and limits: Configuration.

Compare with the live session

Write both sessions' own traffic as CSV and lay them side by side (Compare a backtest with the live session):

./build/release/bin/fastmm-data events runs/mm.12.fmj runs/mm.13.fmj --from-ns 1792483200000000000 --out live.csv
./build/release/bin/fastmm-data events runs/replay/session.fmj --out replay.csv

The run's journal replays bit for bit: fastmm-replay --journal runs/replay/session.fmj --verify exits 0.

Compare two configurations from a checkpoint

Add a branch test to the backtest's configuration (Run an A/B test):

[ab]
name = "spread-wide"
assignment = "branch"

[[ab.arms]]
name = "control"

[[ab.arms]]
name = "wide"
params = { half_spread_bps = 8 }

./build/release/bin/fastmm-backtest --config branch.toml --data "journal:runs/mm.12.fmj,strip_own=1" --out runs/branch-1020

Each arm runs from the same checkpoint over the same data and seed. Its results and journal go to runs/branch-1020/<arm>/, followed by the fastmm-data ab readout.

An hour's window turns on a few fills. Run one branch per checkpoint over several days, each with its own start_at and --out, and read them out together; each branch is one cluster of the resampling:

./build/release/bin/fastmm-data ab runs/branch-*/*/session.fmj