Order Flow Imbalance (OFI)

Quantify net buying pressure at the top of the book from quote-level updates. Order Flow Imbalance (OFI), introduced by Cont, Kukanov, and Stoikov (2014), captures the signed contribution of every L1 quote change (bid arrivals, cancellations, and resizing) into a single number per venue or per symbol.

Unlike trade-based imbalance metrics (aggressor volume imbalance, VPIN), OFI is computed from quotes, not trades. It reacts to liquidity providers reshaping the book before any execution happens, making it a leading indicator of short-term price changes on most liquid instruments.

Problem​

You want to measure top-of-book pressure tick by tick, separating buy-side intent (rising bids, growing bid size, ask cancellations) from sell-side intent. Trade-based metrics miss everything that happens between executions. Limit order placement and cancellation at the best price are often the earliest signal of where the market is about to move.

Solution​

For each consecutive pair of quote updates on a given venue, compute the signed change at the best bid and the best ask, then take their difference. Sum the per-event contributions into time buckets.

OFI per ECN, bucketed at 1 secondDemo this query
WITH quote_lag AS (
    SELECT
        timestamp,
        symbol,
        ecn,
        bid_price, bid_volume, ask_price, ask_volume,
        lag(bid_price)  OVER w AS prev_bid_price,
        lag(bid_volume) OVER w AS prev_bid_volume,
        lag(ask_price)  OVER w AS prev_ask_price,
        lag(ask_volume) OVER w AS prev_ask_volume
    FROM core_price
    WHERE symbol = 'EURUSD'
      AND timestamp IN '$yesterday'
    WINDOW w AS (PARTITION BY symbol, ecn ORDER BY timestamp)
),
contributions AS (
    SELECT
        timestamp,
        symbol,
        ecn,
        CASE
            WHEN bid_price > prev_bid_price THEN bid_volume
            WHEN bid_price < prev_bid_price THEN -prev_bid_volume
            ELSE bid_volume - prev_bid_volume
        END
        -
        CASE
            WHEN ask_price < prev_ask_price THEN ask_volume
            WHEN ask_price > prev_ask_price THEN -prev_ask_volume
            ELSE ask_volume - prev_ask_volume
        END AS ofi_event
    FROM quote_lag
    WHERE prev_bid_price IS NOT NULL
)
SELECT
    timestamp,
    symbol,
    ecn,
    sum(ofi_event) AS ofi
FROM contributions
SAMPLE BY 1s;

How it works​

Step 1: Lag each quote against its previous update​

The quote_lag CTE attaches the previous bid/ask price and volume to every row, partitioned by symbol and ecn so each venue is compared against its own history. The named window w is reused across the four lag calls to keep the query readable.

Step 2: Apply the OFI event rule​

Each row contributes a signed value derived from how the best bid and best ask changed since the last quote, following Cont-Kukanov-Stoikov.

Bid side:

  • Bid price went up: a buyer raised the best bid. The full new size enters as positive pressure.
  • Bid price went down: the previous bid was pulled or executed away. The full old size leaves as negative pressure.
  • Bid price unchanged: it is a size update. The contribution is bid_volume - prev_bid_volume.

Ask side mirrors the bid logic with inverted price comparisons. A lower ask is buy pressure (sellers moving toward buyers). A higher ask is sell pressure.

The event-level OFI is bid contribution − ask contribution. Positive means net buying pressure, negative means net selling pressure.

Step 3: Aggregate into time buckets​

SAMPLE BY 1s sums the per-event contributions into one-second buckets. The bucket size is a tuning knob: it should match the horizon over which you want to relate OFI to price moves. Use 100ms (SAMPLE BY 100T) for microstructure work, or longer windows for slower strategies.

Interpreting results​

Sign indicates direction, magnitude indicates intensity. Across a sample of healthy quote flow, OFI should oscillate around zero with the residual being the predictive signal. A persistent one-sided OFI over many buckets indicates a directional regime, either a genuine trend or a market maker stepping back from one side of the book.

OFI typically leads same-bucket mid-price changes on liquid pairs. To validate that your data and parameters produce a usable signal, compute the correlation between OFI and the mid-price change over the same bucket. On real markets this is typically 0.3 to 0.7 for major FX pairs.

Consolidated OFI across venues​

To get a single OFI series across all ECNs, wrap the per-ECN events and sum:

Consolidated OFI across ECNsDemo this query
WITH quote_lag AS (
    SELECT
        timestamp,
        symbol,
        ecn,
        bid_price, bid_volume, ask_price, ask_volume,
        lag(bid_price)  OVER w AS prev_bid_price,
        lag(bid_volume) OVER w AS prev_bid_volume,
        lag(ask_price)  OVER w AS prev_ask_price,
        lag(ask_volume) OVER w AS prev_ask_volume
    FROM core_price
    WHERE symbol = 'EURUSD'
      AND timestamp IN '$yesterday'
    WINDOW w AS (PARTITION BY symbol, ecn ORDER BY timestamp)
),
contributions AS (
    SELECT
        timestamp,
        symbol,
        CASE
            WHEN bid_price > prev_bid_price THEN bid_volume
            WHEN bid_price < prev_bid_price THEN -prev_bid_volume
            ELSE bid_volume - prev_bid_volume
        END
        -
        CASE
            WHEN ask_price < prev_ask_price THEN ask_volume
            WHEN ask_price > prev_ask_price THEN -prev_ask_volume
            ELSE ask_volume - prev_ask_volume
        END AS ofi_event
    FROM quote_lag
    WHERE prev_bid_price IS NOT NULL
)
SELECT timestamp, symbol, sum(ofi_event) AS ofi
FROM contributions
SAMPLE BY 1s;

This treats each ECN's L1 as an independent book and sums their pressures. It is a good proxy for aggregate flow, but it is not the same as OFI on a virtual consolidated top-of-book (which would require building the best bid/ask across venues first, then computing OFI on that series).

Tuning parameters​

  • Bucket size (SAMPLE BY 1s): match this to your prediction horizon. The correlation with mid-change is highest when the OFI bucket and the return horizon are aligned.
  • Symbol filter: OFI is computed per (symbol, ecn). The WHERE clause keeps the example fast on the demo. Drop it to compute OFI across all symbols at once. The PARTITION BY symbol, ecn in the window handles separation.
  • Time window: $yesterday covers a full session. For intraday work, narrow to a specific hour to keep result sizes manageable.
Related documentation